SlideShare a Scribd company logo
1 of 98
Download to read offline
Countering Bias in Personalized Rankings:
From Data Engineering to Algorithm Development
IEEE ICDE 2021
37th IEEE International Conference on Data Engineering
About us
2
Ludovico Boratto
Senior Research Scientist
EURECAT - Centre Tecnológic de Catalunya
Barcelona, Spain
ludovicoboratto.com
ludovico.boratto@acm.org
Mirko Marras
Postdoctoral Researcher
École Polytechnique Fédérale de Lausanne (EPFL)
Lausanne, Switzerland
mirkomarras.com
mirko.marras@acm.org
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Learning objectives
● Raise awareness on the importance and the relevance of considering data and algorithmic bias issues
in recommendation
● Play with recommendation pipelines and conduct exploratory analysis aimed at uncovering sources
of bias along them
● Showcase approaches that mitigate bias along with the recommendation pipeline and assess their
influence on stakeholders
● Provide an overview on the current trends and challenges in bias-aware research and identify new
research directions
3
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Outline and scheduling
● Welcome and Opening Remarks
● Part I: Recommendation Principles [Slides]
● Part II: Data and Algorithmic Bias Fundamentals [Slides]
● Part III: Bias and Mitigation through the Pipeline [Slides]
● Part IV: Hands on Recommender Systems [Notebook ]
● Part V: Hands on Item Popularity Bias [Notebook ]
● Concluding Remarks [Slides]
4
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Part I
Recommendation
principles
What products could I buy?
6
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
What courses could I attend?
7
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
The problem
8
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
Recommender
System
A solution
9
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
Capitalizing on recommender systems
A recommender system suggests items that might be relevant for a user
10
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
The recommendation ranking task
11
● Given:
○ a set of consumers C = {c1
, c2
, ..., cM
}
○ a set of items I = {i1
, i2
, ..., iN
}
● Let R ⊆ R M ×N
be the consumer-item feedback matrix:
○ R(c,i) ≥ 0 if consumer c expressed interest in item i
○ R(c,i) = 0 otherwise
● The objective is to predict unobserved consumer-item feedback R(c,i) = f(c,i | θ) in R:
○ θ denotes model parameters
○ f denotes the function that maps model parameters to the predicted relevance
● Given a consumer c, items not rated by c are ranked by decreasing relevance:
i* = arg max f(c,j | θ)
j ∈ I  Ic
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
Modes of optimization
12
● Pointwise optimization
point-wise approaches take a user-item pair and predict how relevant the item is for that user
● Pairwise optimization
pair-wise approaches digest a triplet of user, observed item, and unobserved item, and minimize the cases when the
unobserved item is more relevant than the observed item for that user
● Listwise optimization
list-wise approaches look at the entire list and build the optimal ordering for that user
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
Core recommendation techniques
Adapted from [Ricci et al. 2015]
13
Technique Background Input Process
Collaborative Ratings from C of items in I Ratings from c of items in I Identify users in C similar to c and
extrapolate from their preferences of i
Content-based Features of items in I Ratings from c of items in I Generate a classifier that fits c's rating
behavior and use it on i
Demographic Demographic information on C and
their ratings of items in I
Demographic information on c Identify users that are demographically
similar to c and extrapolate from their
preferences of i
Utility-based Features of items in I A utility function over items in I that
describes c's preferences
Apply the function to the items and
determine i's rank
Knowledge-based Features of items in I and knowledge of
how these items meet a user's need
A description of c's needs or interests Infer a match between i and c's needs or
interests
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
Core stakeholders in recommendation
[Abdollahpouri et al. 2020]
A recommendation stakeholder is any group or individual that can affect, or is affected by, the delivery of
recommendations to users
14
Consumers Providers System
C P S
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
A sample multi-sided scenario
15
Consumers
Students
Providers
Teachers
System
Online Course Platform
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
Multi-sided recommendation aspects
[Abdollahpouri et al. 2020]
16
Aspect Definition
Multi-stakeholder design A multistakeholder design process is one in which different recommendation stakeholder groups are
identified and consulted in the process of system design
Multi-stakeholder algorithm A multistakeholder recommendation algorithm takes into account the preferences of multiple parties when
generating recommendations, especially when these parties are on different sides of the recommendation
interaction
Multi-stakeholder evaluation A multistakeholder evaluation is one in which the quality of recommendations is assessed across multiple
groups of stakeholders, in addition to a point estimate over the full user population
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Recommendation principles
Part II
Data and algorithmic bias
fundamentals
Motivating example in music
[Mehrotra et al. 2018]
● People frequently listen to music online
● Ratings and frequencies often used to learn patterns
● 1/3 of users listen to at least 20% of unpopular artists
● Why are popular artists favoured?
● Why do users who tend to interact with niche artists
receive the worst recommendations?
18
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals > Motivating examples
Motivating example in education
[Boratto et al. 2019]
● Online course platforms are receiving great attention
● Student's preferences learnt from ratings/enrolments
● The imbalance in popularity among courses
reinforces coverage and concentration biases of
ranked courses
● Popularity bias could impede new courses to emerge
● The market could be dominated by a few teachers
19
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals > Motivating examples
Motivating example in recruiting
[Singh et al. 2018]
20
● Recruiters rely more and more on automated systems
● Based on the job, "best" candidates are suggested
● Small differences in relevance can lead to large
differences in exposure among candidate groups
● Is this winner-take-all allocation of exposure fair, even
if the winner just has a tiny advantage in relevance?
● It might be fairer to distribute exposure proportional
to relevance, even if this leads to a drop in utility
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals > Motivating examples
Disclaimers
21
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● We aim to focus on scientific literature that specifically consider recommender systems
● Pointers to representative scientific events on related concepts applied to ranking systems are given
● References discussed throughout the slides would not be exhaustive
● Refer to the extended bibliography attached to this tutorial for a more comprehensive list of papers
● This tutorial will not cover biases associated to sensitive attributes (unfairness), which have a
dedicated ICDE tutorial
Data and algorithmic bias fundamentals > Motivating examples
Related scientific venues and
initiatives
22
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Scientific tutorials
including:
Dedicated workshops or tracks
including:
Papers in top-tier conferences
including:
● RecSys
● SIGIR
● The Web Conf
● TREC
● UMAP
● WSDM
● CIKM
● ECIR
● KDD
● FAccTRec @ RecSys 2018-2020
● RMSE & Impact RS @ RecSys 2019
● FACTS-IR @ SIGIR 2019
● FATES @ The Web Conf 2019-21
● FAIR-TREC @ TREC 2020-21
● FairUMAP @ UMAP 2018-20
● DAB @ CIKM 2017-19
● Bias @ ECIR 2020-21
● FAT-ML @ ICML 2014-2019
● Fairness and Discrimination in Retrieval and Recommendation @ SIGIR 2019 & RecSys 2019
● Learning to Rank in theory and practice: From Gradient Boosting to Neural Networks and Unbiased Learning @ SIGIR 2019
● Multi-stakeholder Recommendations: Case Studies, Methods and Challenges @ RecSys 2019
● Experimentation with fairness-aware recommendation using librec-auto @ FAT 2020
● Hands-on on Data and Algorithmic Bias in Recommender Systems @ UMAP 2020
● Bias in Personalized Rankings: Concepts to Code @ ICDM 2020
Data and algorithmic bias fundamentals
Special issues in journals
including:
● Special Issue on Fair, Accountable, and Transparent Recommender Systems @ UMUAI
● Special Issue on Algorithmic Bias and Fairness in Search and Recommendation @ IPM
Perspectives impacted by bias
23
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Economics
Law
Rights
Security
Social
Dynamics
bias can introduce
disparate impacts
among providers,
influencing future success
and revenues
bias can affect core user's
rights that are regulated
by law, such as fairness
and discrimination
bias can reinforce
discrimination of certain
user's groups, including
ageism, sexism,
homophobia
bias can lead certain
groups of users or an
entire system to be more
vulnerable to attacks
(e.g., bribery)
bias can influence how
labour markets evolve and
can be amplified as the
technology progresses
Data and algorithmic bias fundamentals > Shaping the context
How biases can influence economy
[Liu et al. 2019b]
● Loan recommender systems are designed to assist
lenders in looking for promising borrowers
● Such systems model lenders’ historical behaviors and
generate personalized recommendations
● Certain geographical regions such as Asia and Africa
dominate recommendation
● Others like Oceania and Eastern Europe barely
receive recommendations
24
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals > Shaping the context
Biases that impact on law
[Tolan et al. 2019]
The right to non-discrimination, which can be undermined by inherent biases, is embedded in the normative
framework of the European Union, e.g.:
● Mentions can be found in Art. 21 of the EU Charter of Fundamental Rights
● Article 14 of the European Convention on Human Rights
● Articles 18-25 of the Treaty on the Functioning of the European Union
25
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals > Shaping the context
Biases that impact on rights
[Yao et al. 2017]
As an example, United Nations Sustainable Development Goal 4 aims also to ensure equitable quality
education for all, but Yao et al. observed that:
● In 2010, women accounted for only 18% of the BSc degrees in Computer Science
● Historical rating data of Computer Science courses to be dominated by men
● The model may underestimate women’s preferences and be biased towards men
26
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals > Shaping the context
● Sellers might attack the system by introducing a bias in the ratings
● The attack goal is to bribe users to increase ratings and push recommendations
● A novel Hybrid KNN CF is introduced to deal with this phenomenon
● By means of this novel algorithm:
○ The profitability associated to increasing the ratings is strongly reduced w.r.t. SVD
○ Downgrading the ratings of competitors is not profitable with this approach
○ System is more robust to attacks and more trustable by the users
Biases can affect security aspects
[Ramos et al. 2020]
27
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals > Shaping the context
Biases associated to social dynamics
[Fabbri et al. 2020]
28
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● People recommendation in social networks, with users
divided into groups based on gender
● A range of state-of-the-art algorithms, such as Adamic-Adar,
SALSA, and ALS are inspected
● People recommenders produce disparate visibility on the two
subgroups of gender
● Homophily plays a key role in promoting or reducing visibility
for different subgroups
Lorenz Curves (inequality). Recommendations
introduce more inequality than degree distribution,
and this inequality is stronger in the minority class.
Data and algorithmic bias fundamentals > Shaping the context
Ethical aspects influenced by bias
[Bozdag 2013, Milano et al. 2020]
29
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Moral
recommendation of
inappropriate content
Opacity
black-box algorithms, uninformative
explanations, feedback effects
Privacy
unauthorised data collection, data
leaks, unauthorised inferences
Fairness
observation bias,
population imbalance
Autonomy and Identity
behavioural traps and encroachment
on sense of personal autonomy
Social Exposure
lack of exposure to contrasting
viewpoints, feedback effects
Data and algorithmic bias fundamentals > Shaping the context
On biases against immoral content
[Pantakar et al. 2019]
30
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● News recommender system generates awareness on biased
news, to possibly avoid fake and politically polarized news
● They propose a news clustering and bias score attached to
each news. Recommendation of similar, unbiased content
● With a live-user evaluation, the rankings generated by the
algorithm match with the ones the users would generate
Data and algorithmic bias fundamentals > Shaping the context
Bias affecting users' privacy
[Resheff et al. 2018]
31
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● User representations may be used to recover private user information such
as gender and age, undermining users' privacy
● A privacy-adversarial framework is proposed to eliminate leakage of private
information: an adversarial component is appended to the model for each of
the demographic variables we want to obfuscate, so that the user
representations are optimized to preclude predicting the variables
● Privacy preserving recommendations, minimal overall adverse effect on
recommender performance, fairness of results (all knowledge of the
attributes is scrubbed from the representations used by the model)
Data and algorithmic bias fundamentals > Shaping the context
Influence of bias on autonomy
[Arnold et al. 2018]
32
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● Text recommender systems that support creative tasks (writing
reviews). Do they exhibit unintentional biases in the support that
they offer?
● Contextual recommendations are proposed: (1) selection of the
three most likely next words, (2) generation of the most likely
phrase continuation for each word
● People who get recommended phrasal text entry shortcuts that
are skewed positive, write more positive reviews than when
presented with negative-skewed shortcuts
Data and algorithmic bias fundamentals > Shaping the context
How biases can introduce opacity
[Eslami et al. 2020]
33
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● Opaque algorithms sometimes make biased or deceptive
decisions, many have called for increased transparency
● They conducted an analysis of 242 users’ online discussions
about the Yelp review filtering algorithm
● Users defend this algorithm and its opacity depending on their
engagement with and gain from the algorithm
● Then, adding transparency into the algorithm changed users’
attitudes towards the algorithm
Data and algorithmic bias fundamentals > Shaping the context
(a) A filtered review is presented as “recommended”
to the user who wrote it; (b) This review, however,
presented for other users as a filtered review.
Biased effects on social exposure
[Papakyriakopoulos et al. 2020]
34
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● A segment of the political discussions on social networks
is shaped by users that are over-proportionally active
● By applying geometric topic modeling on German political
comments, they demonstrate that hyperactive users have
a significant role in the political discourse
● By training recommender systems, they illustrate that
models provide very different suggestions to users, when
accounting for or ignoring hyperactive behavior
Data and algorithmic bias fundamentals > Shaping the context
System objectives influenced by bias
[Kaminskas et al. 2017, Namatzadeh et al 2018, Singh et al. 2018]
35
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Utility
Recommendation
Objectives
Novelty
Diversity
Coverage
Serendipity
the degree to which recommended
items are potentially useful and of
interest for the user
the degree of attention
received by (groups of)
items or providers
the degree to which the list has
valuable items not looked for and
generate surprise for the user
the degree to which the generated
recommendations cover the catalog
of available items
the degree to which the list of
retrieved items covers a broad area
of the information space
the degree to which items are unknown
by the user and/or are different from
what the user has seen before
Visibility &
Exposure
Data and algorithmic bias fundamentals > Shaping the context
Impact on utility (and trade-offs)
[Fu et al. 2020]
36
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● Studying recommendation performance according to
the level of activity of users:
○ Inactive users are more susceptible to unsatisfactory
recommendations (insufficient training data)
○ Recommendations are biased by the training records of
more active users
● They proposed an explainable CF + re-ranking to
balance predictions and group/individual fairness
● The disparity in utility is reduced while preserving
recommendation quality
Data and algorithmic bias fundamentals > Shaping the context
Impact on coverage (and trade-offs)
[Dean et al. 2020]
37
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● The amount of recourse available to a user is the
percentage of unseen items that are reachable
● The availability of items in a recommender system is
the percentage of items reachable by a user
● Study on linear preference models (SLIM and MF)
● Unavailable items are less popular than available items
● Users with smaller history have more available
recourse
Data and algorithmic bias fundamentals > Shaping the context
Impact on diversity (and trade-offs)
[Lee and Hosanagar 2019]
● They investigated the impact of collaborative
recommender algorithms commonly used in
e-commerce on sales diversity
● The use of traditional collaborative filters is
associated with a decrease in sales diversity relative
to a world without product recommendations
● The decrease in aggregate sales diversity may not
always be accompanied by a corresponding decrease
in individual-level consumption diversity
38
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals > Shaping the context
Impact on novelty (and trade-offs)
[Boratto et al. 2021]
39
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals > Shaping the context
● They show that the probability of being recommended and
the item true positive rate are biased against the item
popularity
● They propose an in-processing approach aimed at
minimizing the biased correlation between user-item
relevance and item popularity
● With small losses in accuracy, their popularity-mitigation
approach leads to important gains in beyond-accuracy
recommendation quality, especially novelty
Impact on providers' exposure
[Boratto et al. 2020]
40
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● Certain minority groups of providers
are being disproportionately affected
by unintentional discrimination
● They show how adding observation
upsampling can dramatically change
the exposure given to providers
● The ranked lists then provide fairer
exposure, wider minority-group item
coverage, and limited loss in utility
Data and algorithmic bias fundamentals > Shaping the context
Part III
Bias and mitigation
through the pipeline
Recommendation pipeline
42
Platform
Data
Model
Recommendations
Data
Preparation
Model
Prediction
Recommendation
Delivering
Model
Evaluation
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Pre-
Processed
Data
Model Setup
and Training
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Data Acquisition
and Storage
Recommendation pipeline: users
43
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Types of bias associated to users
[Olteanu et al. 2017]
● Population biases
differences in demographics between a population of users represented in a dataset/platform and a target population
● Behavioral biases
differences in user behavior across platforms or contexts, or across users represented in different datasets
● Content biases
behavioral biases that are expressed as lexical, syntactic, semantic, and structural differences in the contents generated by users
● Linking biases
behavioral biases expressed as differences in the attributes of networks obtained from user connections, interactions or activity
● Temporal biases
differences in populations or behaviors over time
44
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Bias on items due to their popularity
[Jannach et al. 2014]
● Context: movies, books, hotels, and mobile games
● Algorithms: CB-Filtering, SlopeOne, User-KNN, Item-KNN, FM, RfRec, Funk-SVD, Koren-MF, ALS,
BPR
● Findings: techniques performing well on accuracy focus their recommendations on a tiny fraction of
the item spectrum or recommend mostly top sellers
45
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
On behavioral bias towards popularity
[Cañamares and Castells 2018]
● Context: movies
● Algorithms: User KNN, Item KNN, AvgRating, MF, Random, Pop
● Findings: effectiveness or ineffectiveness of popularity depends on the interplay of three main
variables: item relevance, item discovery by users, and the decision by users to interact with
discovered items. Authors identify the key probabilistic dependencies among these factors
46
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Behavioral patterns on item categories
[Guo and Dunson 2015, Lin et al. 2019]
● Items of different genres have different rating values and different samples
● Bayesian multiplicative probit model to uncover category-wise bias in ratings
47
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● User preferences propagate differently in recommendations,
according to movie genre and user gender
● SVD++ and BiasedMF dampen the preference bias for movie
genres for both men and women
● WRMF is well-calibrated for Sci-Fi/Crime for both men and
women but the behavior is inconsistent for Action/Romance
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Biases conveyed by users' content
[Piramuthu et al. 2012, Xu et al. 2018, Dai et al. 2018, Vall et al. 2019]
● Sequential bias: the sequence in which reviews are written play an
appreciable role in how the next reviews are written
● Opinion bias: given a user–item pair, the opinion bias is defined as the
bias between rating and review. The rating matrix is filled with a
linear combination of the rating and the review sentiment
● Textual bias inspects how recommenders are influenced by the fact
that words may express different meanings in review context
48
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Recommendation pipeline: platform
49
Platform
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Types of bias in platforms
[Olteanu et al. 2017]
● Functional biases
biases resulting from platform-specific mechanisms or affordances, that is, the possible actions within each system or environment
● Normative biases
biases that are a result of written norms or expectations about unwritten norms describing acceptable patterns of behavior on a
given platform
● External biases
biases resulting from factors outside the platform, including considerations of socioeconomic status, education, social pressure,
privacy concerns, interests, language, personality, and culture
● Non-individual accounts
interactions on social platforms that are not produced by individuals, but by accounts representing various types of organizations,
or by automated agents
50
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Functional bias in feedback loop
[Hofmann et al. 2014]
● Context: how a recommender systems evaluation
based on implicit feedback relates to rating-based
evaluation, and how evaluation outcomes may be
affected by bias in user behavior
● Findings:
○ implicit and explicit evaluation agree well when
assumptions agree well (e.g., precision@10)
○ match between assumption on user behavior and
explicit evaluation matters – if assumptions are
violated, the wrong recommender can be preferred
51
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
● Context: characterize the impact of human-system feedback loop in the context of recommender
systems, demonstrating the unintended consequences of algorithmic confounding
● Findings:
○ the recommendation feedback loop
causes homogenization of user behavior
○ users experience losses in utility due
to homogenization effects
○ the feedback loop amplifies the impact
of recommender systems on the
distribution of item consumption
Impact on homogeneity in the loop
[Chaney et al. 2018]
52
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Functional aspects influence choices
[Adomavicius et al 2013]
● Context: explore how consumer preferences are impacted by predictions of recommender systems
● Findings:
○ the rating presented by a recommender serves as an anchor for the consumer’s preference
○ viewers’ preference ratings can be significantly influenced by the recommendation received
○ the effect is sensitive to the perceived reliability of a recommender system
○ the effect of anchoring is continuous and linear, operating over a range of system perturbations
53
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Missing data biases internal functions
[Pradel et al. 2012]
● Context: study two major biases of the selection of
items, i.e., some items obtain more ratings than others
(popularity) and positive ratings are observed more
frequently than negative ratings (positivity)
● Findings:
○ considering missing data as a form of negative feedback
during training may improve performances
○ ...but it can be misleading when testing, favoring
popularity more than user preferences
54
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Deceptive expectations bias user views
[Elsweiler et al. 2017]
● Context: they explore the feasibility of substituting
meals that would typically be recommended to users
with similar, healthier dishes, investigating how
people perceive and select recipes
● Findings:
○ participants are unable to reliably identify which recipe
contains most fat due to their answers being biased by
lack of information
○ perception of fat content can be influenced by the
information available and, in some cases, misleading
cues (image or title) can bias a false impression
55
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
On decision biases in user's choices
[Teppan and Zanker 2015]
● Context: experimental analysis of the impact of different decision biases like decoy or position effects,
as well as risk aversion in positive decision frames
● Findings: risk aversion can be observed in all settings, while position and decoy effects only play a role
when risk aversion is not too predominant
56
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Data Acquisition
and Storage
Recommendation pipeline: data
57
Data
Data
Preparation
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Pre-
Processed
Data
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Sources of bias in data collection
[Olteanu et al. 2017]
● Data acquisition
○ discouraging data collection by third parties
○ programmatic limitation of access to data (e.g., time, amount, size)
○ not all relevant data captured by the platform or opaque and unclear sampling strategies
● Data querying
○ limited expressiveness of APIs regarding information needs
○ different ways of operationalization of information by APIs
○ influence of keywords on datasets in keyword-based queries
● Data filtering
○ removal of outliers that are relevant for the analysis
○ bounding of analysis due to text filtering operations
58
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Sources of bias in data preparation
[Olteanu et al. 2017]
● Data cleaning
○ data representation choices and default values
○ normalization procedures (e.g., based on geographical information)
● Data enrichment
○ subjective and noisy labels due to manual annotations
○ errors due to automatic annotation based on statistical or machine learning
● Data aggregation
○ lose of information due to high-level aggregation
○ spurious patterns of association when data is groups based on certain attributes
59
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Recommendation pipeline: model
60
Model
Recommendations
Model Prediction
Recommendation
Delivering
Model
Evaluation
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Model Setup
and Training
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Sources of bias in model exploitation
[Olteanu et al. 2017]
● Qualitative analysis
○ data representation choices and default values
○ normalization procedures (e.g., based on geographical information)
● Descriptive analysis
○ research often relying on counting entities or influence of bias and confounders on correlation analysis
● Inference analysis
○ definition of target variables, class labels, or data representations
○ effect of the objective function to the inference task
● Observational analysis
○ peer effects due to platform affordances and conventions
○ selection bias and how treatment effects on results generalizability
61
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Sources of bias in model evaluation
[Olteanu et al. 2017, Bellogín et al. 2017]
● Evaluation data selection
○ imbalances of data samples due to their popularity
○ sensitivity to the ratio of the test ratings versus the added non-relevant items
● Metrics selection
○ influence of choice of metrics on research study takeaways
○ accounting domain impact throughout performance assessment
● Result assessment and interpretation
○ traces and patterns changing with context
○ going beyond studies evaluated on a single dataset or method
62
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Bias on popularity in evaluation
[Bellogín et al. 2017]
63
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● First percentile-based approach:
○ dividing items in m popularity percentiles
○ breaking down the computation of accuracy
by such percentiles
○ averaging the m obtained values
● Second uniform test approach:
○ formation of data splits where all items have
the same amount of test ratings
○ picking a set T of candidate items and a
number g of test ratings per item
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Bias and random decoys in evaluation
[Ekstrand et al. 2017] [Other readings: Lim et al. 2015, Yang et al. 2018, Carraro et al. 2020]
● Context: examine the random decoys protocol, where the
candidate set consists of the test set items plus a
randomly-selected set of N decoy items
● Findings:
○ the distribution of items goodness required to avoid
misclassified decoys with reasonable probability is
unreasonable
○ there is a serious discrepancy between theoretical
and observed behavior of the random decoy
strategy with respect to popularity bias
64
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Bias on error estimation in evaluation
[Tian et al. 2020]
● Context: offline evaluation cannot accurately
assess novel, relevant recommendations; the
most novel items are missing from the data
and cannot be judged as relevant
● Findings:
○ missing data in the observation process causes
the evaluation to mis-estimate metric values
○ substantial breakthroughs in recommendation
quality will be difficult to be assessed offline
65
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias through the pipeline
Bias-aware process pipeline
66
IDENTIFY PRODUCT GOALS
● What are you trying to achieve?
● For what population of people?
● What metrics are you tacking?
MITIGATE ISSUES
● Does data include enough minority samples?
● Do our proxies measure what we think they do?
● Does the bias notion capture stakeholders’ needs?
IDENTIFY STAKEHOLDERS
● Who has a stake in this product?
● Who might be harmed?
● How?
DEVELOP AND ANALYZE THE SYSTEM
● How well the system matches product goals?
● To what degree bias is still present?
● How decisions impact on each stakeholder?
DEFINE A BIAS NOTION
● What type of bias? At what point?
● What distributions?
Bias-aware
process pipeline
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Techniques for bias treatment
67
Pre-processing
before model training
In-processing
during model training
Post-processing
after model training
Pre-processing techniques try to
transform the data so that the
bias is mitigated. If the algorithm
is allowed to modify the training
data, pre-processing can be used
In-processing techniques try to
modify learning algorithms to
mitigate bias during training process.
If it is allowed to change the learning
procedure, in-processing can be used
Post-processing is performed by
re-ranking items of the lists obtained
after model training. If the algorithm
can treat the learned model as a black
box, post-processing can be used
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Example of pre-processing treatment
[Jannach et al. 2015]
● Idea: extending BPR with a modified
distribution function φ that samples tuples (u,i, j)
where i is less popular and j is more popular
● Algorithms: BPR
● Findings: it is possible to mitigate popularity
bias, with minor losses in effectiveness
68
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Example of in-processing treatment
[Abdollahpouri et al. 2017]
● Idea: identify a regularization component of the
objective to be minimized when the distribution
of recommendations achieves a 50/50 balance
between medium-tail and short-head items.
● Algorithms: RankALS (i.e., pair-wise learning)
● Findings: it is possible to model the trade-off
between long-tail catalog coverage and ranking
accuracy as a multi-objective optimization
problem based on a dissimilarity matrix
69
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Example of post-processing treatment
[Abdollahpouri et al. 2019]
● Idea: they modified the xQuAD to produce a new re-ranked list
S (|S | < |R|) that manages popularity bias while still being
accurate
● Algorithms: RankALS (i.e., pair-wise learning)
● Findings: on two datasets, the re-ranking methods boost
long-tail items while keeping the accuracy loss small, compared
to the model-based technique
70
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Other treatments for popularity bias
among others...
● Treatments that manipulate interactions before training a model:
○ sample tuples (u,i, j) where i is less popular than j for pair-wise learning [Jannach et al. 2014]
○ remove popular items, simulating situations in which these items are missing [Cremonesi et al. 2014]
○ detect and fix noisy ratings by characterizing items and users by their profiles [Toledo et al. 2015]
● Treatments that regularize the loss function score during training:
○ a regularization that balance recommendation accuracy and intra-list diversity [Abdollahpouri et al. 2017]
○ a regularization that minimizes the correlation between accuracy and item popularity [Boratto et al. 2020a]
○ adversarial framework: minimax game between the BPR model and a discriminator [Zhu et al. 2020]
● Treatments that re-rank items after model training:
○ two-way aggregation of direct and reversed rank results (to improve coverage and accuracy) [Dong et al. 2020]
○ a re-ranking that suggests first items from unseen providers (to improve coverage) [Burke et al. 2016]
○ a re-ranking score that balances predicted rating with the inverse of popularity [Abdollahpouri et al. 2018]
○ a re-ranking that includes long-tail items the user might like [Abdollahpouri et al. 2019]
71
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Treatments against other biases (1)
● Biases related to how items are sampled, positioned, and/or selected, e.g.:
○ connect recommendation to causal inference from experimental and observational data [Schnabel et al. 2016]
○ integrate imputed errors and propensities, for alleviating the effect of propensity variance [Wang et al. 2019]
○ spiral of silence effect, i.e., users are likely to rate if they perceive a support by the dominant opinion [Liu D. et al. 2019a]
○ estimate item frequency from a data stream, subject to vocabulary and distribution shifts [Yi et al. 2019]
○ model position-bias offline and conduct online inference without position information [Guo et al. 2019]
○ off-policy correction to learn from feedback given by an ensemble of prior model policies [Chen et al. 2019a]
○ a clipped estimator to improve the bias-variance trade-off than w.r.t. unbiased estimator [Saito et al. 2020]
○ a counterfactual approach which accounts for selection and position bias jointly [Ovaisi et al. 2020]
○ a two-stage off-policy that takes the model into account while training the candidate model [Ma et al. 2020]
72
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Treatments against other biases (2)
● Biases associated to how items reach their audience:
○ a novel probabilistic method for weighted sampling of k neighbors that considers the similarity levels
between the target user (or item) and the candidate neighbors [Adamopoulos et al. 2014]
○ a target customer re-ranking algorithm to adjust the population distribution and composition in the
top-k target customers of an item while maintaining recommendation quality [Zhao et al. 2020]
● Biases associated to how items are marketed, e.g.:
○ a fairness-aware framework to address market imbalance bias by calibrating the parity of prediction
errors across different market segments [Wan et al. 2020]
73
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Treatments against other biases (3)
● Biases associated to reviews and textual opinions, e.g.:
○ a sentiment classification scoring method, which employs dual attention vectors to predict the users’
sentiment scores of their reviews , to catch opinion bias and enhance user-item matrix [Xu et al. 2018]
○ a hybrid model that integrates modified-sied information related to textual bias and rating bias in matrix
factorization, getting a specific word representation for each item review [Dai et al. 2018]
74
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Treatments against other biases (4)
● Biases associated to social trust and influence:
○ a mitigation using polynomial regression and a Bayesian information criterion to predict ratings less influenced
by the tendency to conform to the perceived “norm” in a community [Krishnan et al. 2014]
○ clustering user-item space to discover rating bubbles derived from the theory of social bias, i.e., existing ratings
indirectly influences the users' opinion to follow the herd instinct [Divyaa et al. 2019]
○ a matrix completion algorithm that performs hybrid memory-based collaborative filtering, improving how the
bribery effect is managed and how the system is robust against bribery [Ramos et al. 2020b]
75
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Treatments against other biases (5)
● Biases related to the interactions of users over time, e.g.:
○ an historical influence-aware latent factor model to capture and mitigate historical distortions in each single
rating under the assimilation-contrast theory: users conform to historical ratings if historical ratings are not far
from the product quality (assimilation), while users deviate from historical ratings if historical ratings are
significantly different from the product quality (contrast) [Zhang et al. 2018]
○ an unbiased loss using inverse propensity weighting, that includes the recency propensity of item x at time t, to be
used in point-wise learning to rank [Chen et al. 2019b]
76
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals
Data and algorithmic bias fundamentals > Bias Mitigation
Part IV
Hands on
recommender systems
Steps of this hands on
78
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Hands on recommender systems
https://colab.research.google.com/github/biasinrecsys/icde2021/blob/master/notebooks/model_setup.ipynb
1
Data Load
We load data from publicly
available datasets, specifically
focusing on Movielens 1M
(movies)
Data Pre-Processing
We process data to be fed
into the model and we prepare
training based on point- and
pair-wise methods
2
Model Definition and Train
We define the architecture of
the model, setup the training
parameters and run the model
training process
3
Relevance Computation
Given a pre-trained model, we
compute the user-item
relevance scores across all
the user-item pairs
4
Model Evaluation
We compute accuracy and
beyond-accuracy metrics,
such as coverage, novelty, and
diversity
5
Part V
Hands on
item popularity bias
Steps of this hands on
80
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Investigation on item popularity bias
https://colab.research.google.com/github/biasinrecsys/icde2021/blob/master/notebooks/item_popularity_bias.ipynb
1
Model Exploration
We consider on data and
models introduced in the first
hands on to inspect of
popularity impacts on visibility
and exposure of items
Mitigation Setup
We arrange a representative
set of mitigation strategies
against popularity bias in pre-,
in- and post-processing
2
Mitigation Running
We run the mitigation
procedure, inspecting how the
optimization processes
influences popularity values
3
Model Re-Evaluation
We re-run the evaluation of
the first hands on to highlight
how disparities among
popular and unpopular items
are reduced
4
Impact Assessment
We interpret the results
obtained during evaluation in
order to envision how
stakeholders are impacted
5
Concluding Remarks
Contextual challenges
82
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● Stakeholders have different (and conflicting) needs. How can recommender systems account for them?
● Multi-disciplinary approaches to go beyond algorithms (e.g., to link justice and fairness)
● Synthesizing a definition of bias or fairness is challenging
● Creating a common vocabulary to recognize different types of bias and unfairness
● Data to characterize bias phenomena with enough depth is lacking (especially for sensitive attributes)
● There are forms of bias on the Web that have not been studied in the recommendation literature
Concluding Remarks
Operational challenges
83
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
● Measuring and operationalizing bias or fairness. How can we optimize a recommender system for it?
● Can we mitigate multiple forms of bias at the same time?
● Slight changes throughout the pipeline can make a huge difference on impact
● Research and development should be more focused on the real world application
● When mitigating bias we usually trade for other qualities.
● How can we mitigate bias without compromising recommendation quality?
Concluding Remarks
Evaluation challenges
● What if we do not have the sensitive attributes in the collected data?
● How should we select an approach with respect to another (e.g., equity vs equality)?
● How to identify harms in the considered context?
● Will the chosen offline metrics and experiments lead to the desired results online?
● How to inspect whether data generation and collection methods are appropriate?
● How could we take into account both bias goals and efficiency in the real world?
84
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Concluding Remarks
Resources from this tutorial
1. Tutorial website https://biasinrecsys.github.io/icde2021/
2. Github repository https://github.com/biasinrecsys/icde2021
3. Jupyter notebooks
3.1. https://colab.research.google.com/github/biasinrecsys/icde2021/blob/master/notebooks/model_setup.ipynb
3.2. https://colab.research.google.com/github/biasinrecsys/icde2021/blob/master/notebooks/item_popularity_bias.ipynb
85
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
Thank you!
References #1
1. Himan Abdollahpouri, Gediminas Adomavicius, Robin Burke, Ido Guy, Dietmar Jannach, Toshihiro Kamishima, Jan Krasnodebski, Luiz Augusto Pizzato: Multi
Stakeholder recommendation: Survey and research directions. User Model. User Adapt. Interact. 30(1): 127-158 (2020).
2. Himan Abdollahpouri, Robin Burke, Bamshad Mobasher: Managing Popularity Bias in Recommender Systems with Personalized Re-Ranking. FLAIRS Conference
2019: 413-418 (2019).
3. Himan Abdollahpouri, Robin Burke, Bamshad Mobasher. Popularity-Aware Item Weighting for Long-Tail Recommendation. arXiv preprint arXiv:1802.05382 (2018).
4. Himan Abdollahpouri, Robin Burke, Bamshad Mobasher: Controlling Popularity Bias in Learning-to-Rank Recommendation. RecSys 2017: 42-46 (2017).
5. Panagiotis Adamopoulos, Alexander Tuzhilin: On over-specialization and concentration bias of recommendations: probabilistic neighborhood selection in
collaborative filtering systems. RecSys 2014: 153-160 (2014).
6. Gediminas Adomavicius, Jesse C. Bockstedt, Shawn P. Curley, Jingjing Zhang: Do Recommender Systems Manipulate Consumer Preferences? A Study of Anchoring
Effects. Inf. Syst. Res. 24(4): 956-975 (2013).
7. Vito Walter Anelli, Tommaso Di Noia, Eugenio Di Sciascio, Azzurra Ragone, Joseph Trotta: Local Popularity and Time in top-N Recommendation. ECIR (1) 2019:
861-868 (2019).
8. Kenneth C. Arnold, Krysta Chauncey, Krzysztof Z. Gajos: Sentiment Bias in Predictive Text Recommendations Results in Biased Writing. Graphics Interface 2018:
42-49 (2018).
87
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #2
9. Alejandro Bellogín, Pablo Castells, Iván Cantador: Statistical biases in Information Retrieval metrics for recommender systems. Inf. Retr. J. 20(6): 606-634 (2017).
10. Alex Beutel, Jilin Chen, Tulsee Doshi, Hai Qian, Li Wei, Yi Wu, Lukasz Heldt, Zhe Zhao, Lichan Hong, Ed H. Chi, Cristos Goodrow: Fairness in Recommendation
Ranking through Pairwise Comparisons. KDD 2019: 2212-2220
11. Asia J. Biega, Krishna P. Gummadi, Gerhard Weikum: Equity of Attention: Amortizing Individual Fairness in Rankings. SIGIR 2018: 405-414
12. Black, J. S., & van Esch, P. (2020). AI-enabled recruiting: What is it and how should a manager use it?. Business Horizons, 63(2), 215-226.
13. Ludovico Boratto, Gianni Fenu, Mirko Marras: The Effect of Algorithmic Bias on Recommender Systems for Massive Open Online Courses. ECIR (1) 2019: 457-472
(2019).
14. Ludovico Boratto, Gianni Fenu, Mirko Marras: Connecting User and Item Perspectives in Popularity Debiasing for Collaborative Recommendation. CoRR
abs/2006.04275 (2020a).
15. Ludovico Boratto, Gianni Fenu, Mirko Marras: Interplay between Upsampling and Regularization for Provider Fairness in Recommender Systems. CoRR
abs/2006.04279 (2020b).
16. Engin Bozdag: Bias in algorithmic filtering and personalization. Ethics Inf Technol 15, 209–227 (2013).
17. Robin Burke. Multisided fairness for recommendation. arXiv preprint arXiv:1707.00093 (2017).
18. Robin Burke, Nasim Sonboli, Aldo Ordonez-Gauger: Balanced Neighborhoods for Multi-sided Fairness in Recommendation. FAT 2018: 202-214 (2018)
19. Robin D. Burke, Himan Abdollahpouri, Bamshad Mobasher, Trinadh Gupta: Towards Multi-Stakeholder Utility Evaluation of Recommender Systems. UMAP
Extended Proceedings (2016).
88
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #3
20. Rocío Cañamares, Pablo Castells: Should I Follow the Crowd?: A Probabilistic Analysis of the Effectiveness of Popularity in Recommender Systems. SIGIR 2018:
415-424
21. Diego Carraro, Derek Bridge: Debiased offline evaluation of recommender systems: a weighted-sampling approach. SAC 2020: 1435-1442 (2020).
22. L. Elisa Celis, Damian Straszak, Nisheeth K. Vishnoi: Ranking with Fairness Constraints. ICALP 2018: 28:1-28:15 (2017).
23. Roberto Centeno, Ramón Hermoso, Maria Fasli: On the inaccuracy of numerical ratings: dealing with biased opinions in social networks. Inf. Syst. Frontiers 17(4):
809-825 (2015).
24. Allison J. B. Chaney, Brandon M. Stewart, Barbara E. Engelhardt: How algorithmic confounding in recommendation systems increases homogeneity and decreases
utility. RecSys 2018: 224-232 (2018).
25. Sushma Channamsetty, Michael D. Ekstrand: Recommender Response to Diversity and Popularity Bias in User Profiles. FLAIRS Conference 2017: 657-660 (2017).
26. Minmin Chen, Alex Beutel, Paul Covington, Sagar Jain, Francois Belletti, Ed H. Chi: Top-K Off-Policy Correction for a REINFORCE Recommender System. WSDM
2019: 456-464 (2019a).
27. Ruey-Cheng Chen, Qingyao Ai, Gaya Jayasinghe, W. Bruce Croft: Correcting for Recency Bias in Job Recommendation. CIKM 2019: 2185-2188 (2019b).
28. Paolo Cremonesi, Franca Garzotto, Roberto Pagano, Massimo Quadrana: Recommending without short head. WWW (Companion Volume) 2014: 245-246 (2014).
89
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #4
29. Jiao Dai, Mingming Li, Songlin Hu, Jizhong Han: A Hybrid Model Based on the Rating Bias and Textual Bias for Recommender Systems. ICONIP (2) 2018: 203-214
(2018).
30. Sarah Dean, Sarah Rich, Benjamin Recht: Recommendations and user agency: the reachability of collaboratively-filtered information. FAT* 2020: 436-445 (2020).
31. Yashar Deldjoo, Vito Walter Anelli, Hamed Zamani, Alejandro Bellogín, Tommaso DiNoia - A Flexible Framework for Evaluating User and Item Fairness in
Recommender Systems. In User Modeling and User-Adapted Interaction (2020)
32. Fernando Diaz, Bhaskar Mitra, Michael D. Ekstrand, Asia J. Biega, Ben Carterette: Evaluating Stochastic Rankings with Expected Exposure. CoRR abs/2004.13157
(2020).
33. Divyaa L. R., Nargis Pervin: Towards generating scalable personalized recommendations: Integrating social trust, social bias, and geo-spatial clustering. Decis. Support
Syst. 122 (2019).
34. Qiang Dong, Quan Yuan, Yang-Bo Shi: Alleviating the recommendation bias via rank aggregation. Physica A: Statistical Mechanics and its Applications, 534, 122073.
(2019).
35. Bora Edizel, Francesco Bonchi, Sara Hajian, André Panisson, Tamir Tassa: FaiRecSys: mitigating algorithmic bias in recommender systems. Int. J. Data Sci. Anal. 9(2):
197-213 (2020)
36. David Elsweiler, Christoph Trattner, Morgan Harvey: Exploiting Food Choice Biases for Healthier Recipe Recommendation. SIGIR 2017: 575-584 (2017).
37. Michael D. Ekstrand, Mucun Tian, Ion Madrazo Azpiazu, Jennifer D. Ekstrand, Oghenemaro Anuyah, David McNeill, Maria Soledad Pera: All The Cool Kids, How Do
They Fit In?: Popularity and Demographic Biases in Recommender Evaluation and Effectiveness. FAT 2018: 172-186 (2018a).
90
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #5
38. Michael D. Ekstrand, Mucun Tian, Mohammed R. Imran Kazi, Hoda Mehrpouyan, Daniel Kluver: Exploring author gender in book rating and recommendation. RecSys
2018: 242-250 (2018b)
39. Michael D. Ekstrand, Vaibhav Mahant: Sturgeon and the Cool Kids: Problems with Random Decoys for Top-N Recommender Evaluation. FLAIRS Conference 2017:
639-644 (2017).
40. Eslami, M., Vaccaro, K., Lee, M. K., Elazari Bar On, A., Gilbert, E., & Karahalios, K. (2019, May). User attitudes towards algorithmic opacity and transparency in online
reviewing platforms. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (pp. 1-14).
41. Francesco Fabbri, Francesco Bonchi, Ludovico Boratto, Carlos Castillo: The Effect of Homophily on Disparate Visibility of Minorities in People Recommender
Systems. ICWSM 2020: 165-175 (2020).
42. Golnoosh Farnadi, Pigi Kouki, Spencer Thompson, Sriram Srinivasan, Lise Getoor: A Fairness-aware Hybrid Recommender System. CoRR abs/1809.09030 (2018).
43. Zuohui Fu, Yikun Xian, Ruoyuan Gao, Jieyu Zhao, Qiaoying Huang, Yingqiang Ge, Shuyuan Xu, Shijie Geng, Chirag Shah, Yongfeng Zhang, Gerard de Melo:
Fairness-Aware Explainable Recommendation over Knowledge Graphs. CoRR abs/2006.02046 (2020).
44. Huifeng Guo, Jinkai Yu, Qing Liu, Ruiming Tang, Yuzhou Zhang: PAL: a position-bias aware learning framework for CTR prediction in live recommender systems.
RecSys 2019: 452-456 (2019).
45. Fangjian Guo, David B. Dunson: Uncovering Systematic Bias in Ratings across Categories: a Bayesian Approach. RecSys 2015: 317-320 (2015).
46. Katja Hofmann, Anne Schuth, Alejandro Bellogín, Maarten de Rijke: Effects of Position Bias on Click-Based Recommender Evaluation. ECIR 2014: 624-630 (2014).
47. Longke Hu, Aixin Sun, Yong Liu: Your neighbors affect your ratings: on geographical neighborhood influence to rating prediction. SIGIR 2014: 345-354 (2014).
91
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #6
48. Farnaz Jahanbakhsh, Justin Cranshaw, Scott Counts, Walter S. Lasecki, Kori Inkpen: An Experimental Study of Bias in Platform Worker Ratings: The Role of
Performance Quality and Gender. CHI 2020: 1-13 (2020).
49. Dietmar Jannach, Lukas Lerche, Iman Kamehkhosh, Michael Jugovac: What recommenders recommend: an analysis of recommendation biases and possible
countermeasures. User Model. User Adapt. Interact. 25(5): 427-491 (2015).
50. Marius Kaminskas, Derek Bridge: Diversity, Serendipity, Novelty, and Coverage: A Survey and Empirical Analysis of Beyond-Accuracy Objectives in Recommender
Systems. ACM Trans. Interact. Intell. Syst. 7(1): 2:1-2:42 (2017).
51. Toshihiro Kamishima, Shotaro Akaho, Hideki Asoh, Jun Sakuma: Recommendation Independence. FAT 2018: 187-201 (2018).
52. Mesut Kaya, Derek Bridge, and Nava Tintarev. 2020. Ensuring Fairness in Group Recommendations by Rank-Sensitive Balancing of Relevance. In Fourteenth ACM
Conference on Recommender Systems (RecSys '20).
53. Dominik Kowald, Markus Schedl, Elisabeth Lex: The Unfairness of Popularity Bias in Music Recommendation: A Reproducibility Study. ECIR (2) 2020: 35-42 (2020).
54. Sanjay Krishnan, Jay Patel, Michael J. Franklin, Ken Goldberg: A methodology for learning, analyzing, and mitigating social influence bias in recommender systems.
RecSys 2014: 137-144 (2014).
55. Caitlin Kuhlman, MaryAnn Van Valkenburg, Elke A. Rundensteiner: FARE: Diagnostics for Fair Ranking using Pairwise Error Metrics. WWW 2019: 2936-2942 (2019).
56. Preethi Lahoti, Krishna P. Gummadi, Gerhard Weikum: iFair: Learning Individually Fair Data Representations for Algorithmic Decision Making. ICDE 2019:
1334-1345
92
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #7
57. Dokyun Lee, Kartik Hosanagar: How Do Recommender Systems Affect Sales Diversity? A Cross-Category Investigation via Randomized Field Experiment. Inf. Syst.
Res. 30(1): 239-259 (2019).
58. Jurek Leonhardt, Avishek Anand, Megha Khosla: User Fairness in Recommender Systems. WWW (Companion Volume) 2018: 101-102 (2018).
59. Mingming Li, Jiao Dai, Fuqing Zhu, Liangjun Zang, Songlin Hu, Jizhong Han: A Fuzzy Set Based Approach for Rating Bias. AAAI 2019: 9969-9970 (2019).
60. Daryl Lim, Julian J. McAuley, Gert R. G. Lanckriet: Top-N Recommendation with Missing Implicit Feedback. RecSys 2015: 309-312 (2015).
61. Xiao Lin, Min Zhang, Yongfeng Zhang, Zhaoquan Gu, Yiqun Liu, Shaoping Ma: Fairness-Aware Group Recommendation with Pareto-Efficiency. RecSys 2017: 107-115
(2017).
62. Kun Lin, Nasim Sonboli, Bamshad Mobasher, Robin Burke: Crank up the Volume: Preference Bias Amplification in Collaborative Recommendation. RMSE@RecSys
2019 (2019).
63. Dugang Liu, Chen Lin, Zhilin Zhang, Yanghua Xiao, Hanghang Tong: Spiral of Silence in Recommender Systems. WSDM 2019: 222-230 (2019a).
64. Weiwen Liu, Jun Guo, Nasim Sonboli, Robin Burke, Shengyu Zhang: Personalized fairness-aware re-ranking for microlending. RecSys 2019: 467-471 (2019b).
65. Jiaqi Ma, Zhe Zhao, Xinyang Yi, Ji Yang, Minmin Chen, Jiaxi Tang, Lichan Hong, Ed H. Chi: Off-policy Learning in Two-stage Recommender Systems. WWW 2020:
463-473 (2020).
66. Benjamin M. Marlin, Richard S. Zemel, Sam T. Roweis, Malcolm Slaney: Recommender Systems, Missing Data and Statistical Model Estimation. IJCAI 2011:
2686-2691 (2011).
93
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #8
68. Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, Aram Galstyan. A survey on bias and fairness in machine learning. arXiv preprint
arXiv:1908.09635 (2019).
69. Rishabh Mehrotra, James McInerney, Hugues Bouchard, Mounia Lalmas, Fernando Diaz: Towards a Fair Marketplace: Counterfactual Evaluation of the trade-off
between Relevance, Fairness & Satisfaction in Recommendation Systems. CIKM 2018: 2243-2251 (2018).
70. Silvia Milano, Mariarosaria Taddeo, Luciano Floridi. Recommender systems and their ethical challenges. AI & Soc (2020).
71. Natwar Modani, Deepali Jain, Ujjawal Soni, Gaurav Kumar Gupta, Palak Agarwal: Fairness Aware Recommendations on Behance. PAKDD (2) 2017: 144-155 (2017).
72. Azadeh Nematzadeh, Giovanni Luca Ciampaglia, Filippo Menczer, Alessandro Flammini: How algorithmic popularity bias hinders or promotes quality. CoRR
abs/1707.00574 (2017).
73. Alexandra Olteanu, Carlos Castillo, Fernando Diaz, Emre Kiciman: Social Data: Biases, Methodological Pitfalls, and Ethical Boundaries. Frontiers Big Data 2: 13 (2019)
74. Zohreh Ovaisi, Ragib Ahsan, Yifan Zhang, Kathryn Vasilaky, Elena Zheleva: Correcting for Selection Bias in Learning-to-rank Systems. WWW 2020: 1863-1873 (2020).
75. Papakyriakopoulos, Orestis, Juan Carlos Medina Serrano, and Simon Hegelich. "Political communication on social media: A tale of hyperactive users and bias in
recommender systems." Online Social Networks and Media 15 (2020): 100058.
76. Anish Anil Patankar, Joy Bose, Harshit Khanna: A Bias Aware News Recommendation System. ICSC 2019: 232-238 (2019).
77. Bruno Pradel, Nicolas Usunier, Patrick Gallinari: Ranking with non-random missing ratings: influence of popularity and positivity on evaluation metrics. RecSys 2012:
147-154 (2012).
94
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #9
78. Gourab K. Patro, Arpita Biswas, Niloy Ganguly, Krishna P. Gummadi, Abhijnan Chakraborty: FairRec: Two-Sided Fairness for Personalized Recommendations in
Two-Sided Platforms. WWW 2020: 1194-1204 (2020).
79. Gourab K. Patro, Abhijnan Chakraborty, Niloy Ganguly, Krishna Gummadi. Incremental Fairness in Two-Sided Market Platforms: On Smoothly Updating
Recommendations. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 34, No. 01, pp. 181-188).
80. Selwyn Piramuthu, Gaurav Kapoor, Wei Zhou, Sjouke Mauw: Input online review data and related bias in recommender systems. Decis. Sup. Sys. 53(3): 418-424 (2012).
81. Raghavan, M., Barocas, S., Kleinberg, J., & Levy, K. (2020, January). Mitigating bias in algorithmic hiring: Evaluating claims and practices. In Proceedings of the 2020
conference on fairness, accountability, and transparency (pp. 469-481).
82. Guilherme Ramos and Ludovico Boratto, Reputation (In)dependence in Ranking Systems: Demographics Influence Over Output Disparities, in Proceedings of the 43rd
International ACM SIGIR Conference on Researchand Development in Information Retrieval, SIGIR 2020 (2020a)
83. Guilherme Ramos, Ludovico Boratto, Carlos Caleiro. On the negative impact of social influence in recommender systems: A study of bribery in collaborative hybrid
algorithms. Information Processing & Management, 57(2), 102058 (2020).
84. Bashir Rastegarpanah, Krishna P. Gummadi, Mark Crovella: Fighting Fire with Fire: Using Antidote Data to Improve Polarization and Fairness of Recommender Systems.
WSDM 2019: 231-239
85. Yehezkel S. Resheff, Yanai Elazar, Moni Shahar, Oren Sar Shalom: Privacy and Fairness in Recommender Systems via Adversarial Training of User Representations.
ICPRAM 2019: 476-482
86. Francesco Ricci, Lior Rokach, Bracha Shapira: Recommender Systems: Introduction and Challenges. Recommender Systems Handbook 2015: 1-34 (2015).
95
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #10
87. Yuta Saito, Suguru Yaginuma, Yuta Nishino, Hayato Sakata, Kazuhide Nakata: Unbiased Recommender Learning from Missing-Not-At-Random Implicit Feedback.
WSDM 2020: 501-509 (2020).
88. Tobias Schnabel, Adith Swaminathan, Ashudeep Singh, Navin Chandak, Thorsten Joachims: Recommendations as Treatments: Debiasing Learning and Evaluation.
ICML 2016: 1670-1679 (2016).
89. Christina Schwind, Jürgen Buder: Reducing confirmation bias and evaluation bias: When are preference-inconsistent recommendations effective - and when not?
Comput. Hum. Behav. 28(6): 2280-2290 (2012).
90. Dimitris Serbos, Shuyao Qi, Nikos Mamoulis, Evaggelia Pitoura, Panayiotis Tsaparas: Fairness in Package-to-Group Recommendations. WWW 2017: 371-379 (2017).
91. Shakespeare, D., Porcaro, L., Gómez, E., & Castillo, C. Exploring Artist Gender Bias in Music Recommendation.
92. Ashudeep Singh, Thorsten Joachims: Policy Learning for Fairness in Ranking. NeurIPS 2019: 5427-5437 (2019).
93. Ashudeep Singh, Thorsten Joachims: Fairness of Exposure in Rankings. KDD 2018: 2219-2228 (2018).
94. Nasim Sonboli, Robin Burke: Localized Fairness in Recommender Systems. UMAP (Adjunct Publication) 2019: 295-300 (2019).
95. Harald Steck: Item popularity and recommendation accuracy. RecSys 2011: 125-132 (2011).
96. Maria Stratigi, Haridimos Kondylakis, Kostas Stefanidis: Fairness in Group Recommendations in the Health Domain. ICDE 2017: 1481-1488 (2017).
97. Erich Teppan, Marcus Zanker, M: Decision biases in recommender systems. Journal of Internet Commerce, 14(2), 255-275 (2015).
98. Mucun Tian, Michael D. Ekstrand: Estimating Error and Bias in Offline Evaluation Results. CHIIR 2020: 392-396 (2020).
96
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #11
99. Songül Tolan: Fair and Unbiased Algorithmic Decision Making: Current State and Future Challenges. CoRR abs/1901.04730 (2019).
100. Raciel Yera Toledo, Yailé Caballero Mota, Luis Martínez-López: Correcting noisy ratings in collaborative recommender systems. Knowl. Based Syst. 76: 96-108 (2015).
101. Andreu Vall, Massimo Quadrana, Markus Schedl, Gerhard Widmer: Order, context and popularity bias in next-song recommendations. Int. J. Multim. Inf. Retr. 8(2):
101-113 (2019).
102. Mengting Wan, Jianmo Ni, Rishabh Misra, Julian J. McAuley: Addressing Marketing Bias in Product Recommendations. WSDM 2020: 618-626 (2020).
103. Xiaojie Wang, Rui Zhang, Yu Sun, Jianzhong Qi: Doubly Robust Joint Learning for Recommendation on Data Missing Not at Random. ICML 2019: 6638-6647 (2019).
104. Jacek Wasilewski, Neil Hurley: Are You Reaching Your Audience?: Exploring Item Exposure over Consumer Segments in Recommender Systems. UMAP 2018:
213-217 (2018).
105. Bin Xia, Junjie Yin, Jian Xu, Yun Li: WE-Rec: A fairness-aware reciprocal recommendation based on Walrasian equilibrium. Knowl. Based Syst. 182 (2019).
106. Yuanbo Xu, Yongjian Yang, Jiayu Han, En Wang, Fuzhen Zhuang, Hui Xiong: Exploiting the Sentimental Bias between Ratings and Reviews for Enhancing
Recommendation. ICDM 2018: 1356-1361 (2018).
107. Himank Yadav, Zhengxiao Du, Thorsten Joachims: Fair Learning-to-Rank from Implicit Feedback. CoRR abs/1911.08054 (2019).
108. Longqi Yang, Yin Cui, Yuan Xuan, Chenyang Wang, Serge J. Belongie, Deborah Estrin: Unbiased offline recommender evaluation for missing-not-at-random implicit
feedback. RecSys 2018: 279-287 (2018).
97
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras
References #12
110. Ke Yang, Julia Stoyanovich: Measuring Fairness in Ranked Outputs. SSDBM 2017: 22:1-22:6 (2017).
111. Sirui Yao, Bert Huang: Beyond Parity: Fairness Objectives for Collaborative Filtering. NIPS 2017: 2921-2930 (2017).
112. Xinyang Yi, Ji Yang, Lichan Hong, Derek Zhiyuan Cheng, Lukasz Heldt, Aditee Kumthekar, Zhe Zhao, Li Wei, Ed H. Chi: Sampling-bias-corrected neural modeling for
large corpus item recommendations. RecSys 2019: 269-277 (2019).
113. Meike Zehlike, Carlos Castillo: Reducing Disparate Exposure in Ranking: A Learning To Rank Approach. WWW 2020: 2849-2855 (2020).
114. Meike Zehlike, Francesco Bonchi, Carlos Castillo, Sara Hajian, Mohamed Megahed, Ricardo Baeza-Yates: FA*IR: A Fair Top-k Ranking Algorithm. CIKM 2017:
1569-1578 (2017).
115. Shuai Zhang, Lina Yao, Aixin Sun, Yi Tay: Deep Learning Based Recommender System: A Survey and New Perspectives. ACM Comput. Surv. 52(1): 5:1-5:38 (2019)
116. Xiaoying Zhang, Hong Xie, Junzhou Zhao, John C. S. Lui: Modeling the Assimilation-Contrast Effects in Online Product Rating Systems: Debiasing and
Recommendations. IJCAI 2018: 5409-5413 (2018).
117. Xing Zhao, Ziwei Zhu, Majid Alfifi, James Caverlee: Addressing the Target Customer Distortion Problem in Recommender Systems. WWW 2020: 2969-2975 (2020).
118. Yong Zheng, Tanaya Dave, Neha Mishra, Harshit Kumar: Fairness In Reciprocal Recommendations: A Speed-Dating Study. UMAP 2018: 29-34 (2018).
119. Ziwei Zhu, Xia Hu, James Caverlee: Fairness-Aware Tensor-Based Recommendation. CIKM 2018: 1153-1162 (2018).
120. Ziwei Zhu, Jianling Wang, and James Caverlee. 2020. Measuring and Mitigating Item Under-Recommendation Bias in Personalized Ranking Systems. In Proceedings
of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '20).
98
Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development
Boratto and Marras

More Related Content

What's hot

A literature survey on recommendation
A literature survey on recommendationA literature survey on recommendation
A literature survey on recommendationaciijournal
 
Multidirectional Product Support System for Decision Making In Textile Indust...
Multidirectional Product Support System for Decision Making In Textile Indust...Multidirectional Product Support System for Decision Making In Textile Indust...
Multidirectional Product Support System for Decision Making In Textile Indust...IOSR Journals
 
Customer Point of View on Operations Management
Customer Point of View on Operations ManagementCustomer Point of View on Operations Management
Customer Point of View on Operations ManagementBaker Khader Abdallah, PMP
 
Customer point of view on operations management
Customer point of view on operations managementCustomer point of view on operations management
Customer point of view on operations managementBaker Khader Abdallah, PMP
 
Certified Machine Learning Specialist (CMLS)
Certified Machine Learning Specialist (CMLS)Certified Machine Learning Specialist (CMLS)
Certified Machine Learning Specialist (CMLS)GICTTraining
 

What's hot (6)

A literature survey on recommendation
A literature survey on recommendationA literature survey on recommendation
A literature survey on recommendation
 
Multidirectional Product Support System for Decision Making In Textile Indust...
Multidirectional Product Support System for Decision Making In Textile Indust...Multidirectional Product Support System for Decision Making In Textile Indust...
Multidirectional Product Support System for Decision Making In Textile Indust...
 
Tarikua Woldetsadick - LME Unit
Tarikua Woldetsadick - LME UnitTarikua Woldetsadick - LME Unit
Tarikua Woldetsadick - LME Unit
 
Customer Point of View on Operations Management
Customer Point of View on Operations ManagementCustomer Point of View on Operations Management
Customer Point of View on Operations Management
 
Customer point of view on operations management
Customer point of view on operations managementCustomer point of view on operations management
Customer point of view on operations management
 
Certified Machine Learning Specialist (CMLS)
Certified Machine Learning Specialist (CMLS)Certified Machine Learning Specialist (CMLS)
Certified Machine Learning Specialist (CMLS)
 

Similar to Tutorial on Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development @ ICDE 2021

A Business case study on LinkedIn
A Business case study on LinkedInA Business case study on LinkedIn
A Business case study on LinkedInMayank Banerjee
 
Intelligent Career Guidance System.pptx
Intelligent Career Guidance System.pptxIntelligent Career Guidance System.pptx
Intelligent Career Guidance System.pptxAnonymous366406
 
IRJET-Survey on Identification of Top-K Competitors using Data Mining
IRJET-Survey on Identification of Top-K Competitors using Data MiningIRJET-Survey on Identification of Top-K Competitors using Data Mining
IRJET-Survey on Identification of Top-K Competitors using Data MiningIRJET Journal
 
An Najah University IT Market Skill Needs Survey
An Najah University IT Market Skill Needs SurveyAn Najah University IT Market Skill Needs Survey
An Najah University IT Market Skill Needs SurveyLaith Kassis
 
Infrastructure that can stand the test of time | Accenture
Infrastructure that can stand the test of time | AccentureInfrastructure that can stand the test of time | Accenture
Infrastructure that can stand the test of time | Accentureaccenture
 
Rishabh Mehrotra - Recommendations in a Marketplace: Personalizing Explainabl...
Rishabh Mehrotra - Recommendations in a Marketplace: Personalizing Explainabl...Rishabh Mehrotra - Recommendations in a Marketplace: Personalizing Explainabl...
Rishabh Mehrotra - Recommendations in a Marketplace: Personalizing Explainabl...MLconf
 
Algorithmic auditing 1.0
Algorithmic auditing 1.0Algorithmic auditing 1.0
Algorithmic auditing 1.0QuantUniversity
 
CFA-NY Workshop - Final slides
CFA-NY Workshop - Final slidesCFA-NY Workshop - Final slides
CFA-NY Workshop - Final slidesQuantUniversity
 
machine-learning-development-audit-framework-assessment-and-inspection-of-ris...
machine-learning-development-audit-framework-assessment-and-inspection-of-ris...machine-learning-development-audit-framework-assessment-and-inspection-of-ris...
machine-learning-development-audit-framework-assessment-and-inspection-of-ris...Daniel983829
 
IoT Innovation Design Method (Picmet2019 Presentation)
IoT Innovation Design Method (Picmet2019 Presentation)IoT Innovation Design Method (Picmet2019 Presentation)
IoT Innovation Design Method (Picmet2019 Presentation)Naoshi Uchihira
 
RT332: Measuring Progress and Productivity in Model-based Engineering
RT332: Measuring Progress and Productivity in Model-based EngineeringRT332: Measuring Progress and Productivity in Model-based Engineering
RT332: Measuring Progress and Productivity in Model-based EngineeringAVEVA Group plc
 
GECBSP Guide GDSC India ML Study Jam organizer guide modified.pdf
GECBSP Guide GDSC India ML Study Jam organizer guide modified.pdfGECBSP Guide GDSC India ML Study Jam organizer guide modified.pdf
GECBSP Guide GDSC India ML Study Jam organizer guide modified.pdfDomendra Sahu
 
Mining competitors from large unstructured data
Mining competitors from large unstructured dataMining competitors from large unstructured data
Mining competitors from large unstructured dataTejasSalunkhe11
 
Survey on Study Partners Recommendation for Online Courses
Survey on Study Partners Recommendation for Online CoursesSurvey on Study Partners Recommendation for Online Courses
Survey on Study Partners Recommendation for Online CoursesIRJET Journal
 
Algorithmic auditing 1.0
Algorithmic auditing 1.0Algorithmic auditing 1.0
Algorithmic auditing 1.0QuantUniversity
 

Similar to Tutorial on Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development @ ICDE 2021 (20)

A Business case study on LinkedIn
A Business case study on LinkedInA Business case study on LinkedIn
A Business case study on LinkedIn
 
QCon conference 2019
QCon conference 2019QCon conference 2019
QCon conference 2019
 
CSEIT- ALL.pptx
CSEIT- ALL.pptxCSEIT- ALL.pptx
CSEIT- ALL.pptx
 
Intelligent Career Guidance System.pptx
Intelligent Career Guidance System.pptxIntelligent Career Guidance System.pptx
Intelligent Career Guidance System.pptx
 
IRJET-Survey on Identification of Top-K Competitors using Data Mining
IRJET-Survey on Identification of Top-K Competitors using Data MiningIRJET-Survey on Identification of Top-K Competitors using Data Mining
IRJET-Survey on Identification of Top-K Competitors using Data Mining
 
An Najah University IT Market Skill Needs Survey
An Najah University IT Market Skill Needs SurveyAn Najah University IT Market Skill Needs Survey
An Najah University IT Market Skill Needs Survey
 
Infrastructure that can stand the test of time | Accenture
Infrastructure that can stand the test of time | AccentureInfrastructure that can stand the test of time | Accenture
Infrastructure that can stand the test of time | Accenture
 
Rishabh Mehrotra - Recommendations in a Marketplace: Personalizing Explainabl...
Rishabh Mehrotra - Recommendations in a Marketplace: Personalizing Explainabl...Rishabh Mehrotra - Recommendations in a Marketplace: Personalizing Explainabl...
Rishabh Mehrotra - Recommendations in a Marketplace: Personalizing Explainabl...
 
Algorithmic auditing 1.0
Algorithmic auditing 1.0Algorithmic auditing 1.0
Algorithmic auditing 1.0
 
ML master class
ML master classML master class
ML master class
 
CFA-NY Workshop - Final slides
CFA-NY Workshop - Final slidesCFA-NY Workshop - Final slides
CFA-NY Workshop - Final slides
 
machine-learning-development-audit-framework-assessment-and-inspection-of-ris...
machine-learning-development-audit-framework-assessment-and-inspection-of-ris...machine-learning-development-audit-framework-assessment-and-inspection-of-ris...
machine-learning-development-audit-framework-assessment-and-inspection-of-ris...
 
DTT
DTTDTT
DTT
 
IoT Innovation Design Method (Picmet2019 Presentation)
IoT Innovation Design Method (Picmet2019 Presentation)IoT Innovation Design Method (Picmet2019 Presentation)
IoT Innovation Design Method (Picmet2019 Presentation)
 
RT332: Measuring Progress and Productivity in Model-based Engineering
RT332: Measuring Progress and Productivity in Model-based EngineeringRT332: Measuring Progress and Productivity in Model-based Engineering
RT332: Measuring Progress and Productivity in Model-based Engineering
 
GECBSP Guide GDSC India ML Study Jam organizer guide modified.pdf
GECBSP Guide GDSC India ML Study Jam organizer guide modified.pdfGECBSP Guide GDSC India ML Study Jam organizer guide modified.pdf
GECBSP Guide GDSC India ML Study Jam organizer guide modified.pdf
 
Mining competitors from large unstructured data
Mining competitors from large unstructured dataMining competitors from large unstructured data
Mining competitors from large unstructured data
 
Survey on Study Partners Recommendation for Online Courses
Survey on Study Partners Recommendation for Online CoursesSurvey on Study Partners Recommendation for Online Courses
Survey on Study Partners Recommendation for Online Courses
 
Algorithmic auditing 1.0
Algorithmic auditing 1.0Algorithmic auditing 1.0
Algorithmic auditing 1.0
 
master_thesis.pdf
master_thesis.pdfmaster_thesis.pdf
master_thesis.pdf
 

Recently uploaded

AXA XL - Insurer Innovation Award Americas 2024
AXA XL - Insurer Innovation Award Americas 2024AXA XL - Insurer Innovation Award Americas 2024
AXA XL - Insurer Innovation Award Americas 2024The Digital Insurer
 
presentation ICT roal in 21st century education
presentation ICT roal in 21st century educationpresentation ICT roal in 21st century education
presentation ICT roal in 21st century educationjfdjdjcjdnsjd
 
[BuildWithAI] Introduction to Gemini.pdf
[BuildWithAI] Introduction to Gemini.pdf[BuildWithAI] Introduction to Gemini.pdf
[BuildWithAI] Introduction to Gemini.pdfSandro Moreira
 
Emergent Methods: Multi-lingual narrative tracking in the news - real-time ex...
Emergent Methods: Multi-lingual narrative tracking in the news - real-time ex...Emergent Methods: Multi-lingual narrative tracking in the news - real-time ex...
Emergent Methods: Multi-lingual narrative tracking in the news - real-time ex...Zilliz
 
Strategies for Landing an Oracle DBA Job as a Fresher
Strategies for Landing an Oracle DBA Job as a FresherStrategies for Landing an Oracle DBA Job as a Fresher
Strategies for Landing an Oracle DBA Job as a FresherRemote DBA Services
 
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...apidays
 
MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024MIND CTI
 
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...Angeliki Cooney
 
"I see eyes in my soup": How Delivery Hero implemented the safety system for ...
"I see eyes in my soup": How Delivery Hero implemented the safety system for ..."I see eyes in my soup": How Delivery Hero implemented the safety system for ...
"I see eyes in my soup": How Delivery Hero implemented the safety system for ...Zilliz
 
Spring Boot vs Quarkus the ultimate battle - DevoxxUK
Spring Boot vs Quarkus the ultimate battle - DevoxxUKSpring Boot vs Quarkus the ultimate battle - DevoxxUK
Spring Boot vs Quarkus the ultimate battle - DevoxxUKJago de Vreede
 
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost SavingRepurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost SavingEdi Saputra
 
Modular Monolith - a Practical Alternative to Microservices @ Devoxx UK 2024
Modular Monolith - a Practical Alternative to Microservices @ Devoxx UK 2024Modular Monolith - a Practical Alternative to Microservices @ Devoxx UK 2024
Modular Monolith - a Practical Alternative to Microservices @ Devoxx UK 2024Victor Rentea
 
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Victor Rentea
 
DEV meet-up UiPath Document Understanding May 7 2024 Amsterdam
DEV meet-up UiPath Document Understanding May 7 2024 AmsterdamDEV meet-up UiPath Document Understanding May 7 2024 Amsterdam
DEV meet-up UiPath Document Understanding May 7 2024 AmsterdamUiPathCommunity
 
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Orbitshub
 
Ransomware_Q4_2023. The report. [EN].pdf
Ransomware_Q4_2023. The report. [EN].pdfRansomware_Q4_2023. The report. [EN].pdf
Ransomware_Q4_2023. The report. [EN].pdfOverkill Security
 
Cyberprint. Dark Pink Apt Group [EN].pdf
Cyberprint. Dark Pink Apt Group [EN].pdfCyberprint. Dark Pink Apt Group [EN].pdf
Cyberprint. Dark Pink Apt Group [EN].pdfOverkill Security
 
How to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerHow to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerThousandEyes
 
Manulife - Insurer Transformation Award 2024
Manulife - Insurer Transformation Award 2024Manulife - Insurer Transformation Award 2024
Manulife - Insurer Transformation Award 2024The Digital Insurer
 

Recently uploaded (20)

AXA XL - Insurer Innovation Award Americas 2024
AXA XL - Insurer Innovation Award Americas 2024AXA XL - Insurer Innovation Award Americas 2024
AXA XL - Insurer Innovation Award Americas 2024
 
presentation ICT roal in 21st century education
presentation ICT roal in 21st century educationpresentation ICT roal in 21st century education
presentation ICT roal in 21st century education
 
[BuildWithAI] Introduction to Gemini.pdf
[BuildWithAI] Introduction to Gemini.pdf[BuildWithAI] Introduction to Gemini.pdf
[BuildWithAI] Introduction to Gemini.pdf
 
Emergent Methods: Multi-lingual narrative tracking in the news - real-time ex...
Emergent Methods: Multi-lingual narrative tracking in the news - real-time ex...Emergent Methods: Multi-lingual narrative tracking in the news - real-time ex...
Emergent Methods: Multi-lingual narrative tracking in the news - real-time ex...
 
Strategies for Landing an Oracle DBA Job as a Fresher
Strategies for Landing an Oracle DBA Job as a FresherStrategies for Landing an Oracle DBA Job as a Fresher
Strategies for Landing an Oracle DBA Job as a Fresher
 
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
 
MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024
 
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
 
"I see eyes in my soup": How Delivery Hero implemented the safety system for ...
"I see eyes in my soup": How Delivery Hero implemented the safety system for ..."I see eyes in my soup": How Delivery Hero implemented the safety system for ...
"I see eyes in my soup": How Delivery Hero implemented the safety system for ...
 
Spring Boot vs Quarkus the ultimate battle - DevoxxUK
Spring Boot vs Quarkus the ultimate battle - DevoxxUKSpring Boot vs Quarkus the ultimate battle - DevoxxUK
Spring Boot vs Quarkus the ultimate battle - DevoxxUK
 
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost SavingRepurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
 
Modular Monolith - a Practical Alternative to Microservices @ Devoxx UK 2024
Modular Monolith - a Practical Alternative to Microservices @ Devoxx UK 2024Modular Monolith - a Practical Alternative to Microservices @ Devoxx UK 2024
Modular Monolith - a Practical Alternative to Microservices @ Devoxx UK 2024
 
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
 
DEV meet-up UiPath Document Understanding May 7 2024 Amsterdam
DEV meet-up UiPath Document Understanding May 7 2024 AmsterdamDEV meet-up UiPath Document Understanding May 7 2024 Amsterdam
DEV meet-up UiPath Document Understanding May 7 2024 Amsterdam
 
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
 
Ransomware_Q4_2023. The report. [EN].pdf
Ransomware_Q4_2023. The report. [EN].pdfRansomware_Q4_2023. The report. [EN].pdf
Ransomware_Q4_2023. The report. [EN].pdf
 
Cyberprint. Dark Pink Apt Group [EN].pdf
Cyberprint. Dark Pink Apt Group [EN].pdfCyberprint. Dark Pink Apt Group [EN].pdf
Cyberprint. Dark Pink Apt Group [EN].pdf
 
How to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerHow to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected Worker
 
Manulife - Insurer Transformation Award 2024
Manulife - Insurer Transformation Award 2024Manulife - Insurer Transformation Award 2024
Manulife - Insurer Transformation Award 2024
 
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
 

Tutorial on Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development @ ICDE 2021

  • 1. Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development IEEE ICDE 2021 37th IEEE International Conference on Data Engineering
  • 2. About us 2 Ludovico Boratto Senior Research Scientist EURECAT - Centre Tecnológic de Catalunya Barcelona, Spain ludovicoboratto.com ludovico.boratto@acm.org Mirko Marras Postdoctoral Researcher École Polytechnique Fédérale de Lausanne (EPFL) Lausanne, Switzerland mirkomarras.com mirko.marras@acm.org Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 3. Learning objectives ● Raise awareness on the importance and the relevance of considering data and algorithmic bias issues in recommendation ● Play with recommendation pipelines and conduct exploratory analysis aimed at uncovering sources of bias along them ● Showcase approaches that mitigate bias along with the recommendation pipeline and assess their influence on stakeholders ● Provide an overview on the current trends and challenges in bias-aware research and identify new research directions 3 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 4. Outline and scheduling ● Welcome and Opening Remarks ● Part I: Recommendation Principles [Slides] ● Part II: Data and Algorithmic Bias Fundamentals [Slides] ● Part III: Bias and Mitigation through the Pipeline [Slides] ● Part IV: Hands on Recommender Systems [Notebook ] ● Part V: Hands on Item Popularity Bias [Notebook ] ● Concluding Remarks [Slides] 4 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 6. What products could I buy? 6 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 7. What courses could I attend? 7 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 8. The problem 8 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 9. Recommender System A solution 9 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 10. Capitalizing on recommender systems A recommender system suggests items that might be relevant for a user 10 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 11. The recommendation ranking task 11 ● Given: ○ a set of consumers C = {c1 , c2 , ..., cM } ○ a set of items I = {i1 , i2 , ..., iN } ● Let R ⊆ R M ×N be the consumer-item feedback matrix: ○ R(c,i) ≥ 0 if consumer c expressed interest in item i ○ R(c,i) = 0 otherwise ● The objective is to predict unobserved consumer-item feedback R(c,i) = f(c,i | θ) in R: ○ θ denotes model parameters ○ f denotes the function that maps model parameters to the predicted relevance ● Given a consumer c, items not rated by c are ranked by decreasing relevance: i* = arg max f(c,j | θ) j ∈ I Ic Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 12. Modes of optimization 12 ● Pointwise optimization point-wise approaches take a user-item pair and predict how relevant the item is for that user ● Pairwise optimization pair-wise approaches digest a triplet of user, observed item, and unobserved item, and minimize the cases when the unobserved item is more relevant than the observed item for that user ● Listwise optimization list-wise approaches look at the entire list and build the optimal ordering for that user Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 13. Core recommendation techniques Adapted from [Ricci et al. 2015] 13 Technique Background Input Process Collaborative Ratings from C of items in I Ratings from c of items in I Identify users in C similar to c and extrapolate from their preferences of i Content-based Features of items in I Ratings from c of items in I Generate a classifier that fits c's rating behavior and use it on i Demographic Demographic information on C and their ratings of items in I Demographic information on c Identify users that are demographically similar to c and extrapolate from their preferences of i Utility-based Features of items in I A utility function over items in I that describes c's preferences Apply the function to the items and determine i's rank Knowledge-based Features of items in I and knowledge of how these items meet a user's need A description of c's needs or interests Infer a match between i and c's needs or interests Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 14. Core stakeholders in recommendation [Abdollahpouri et al. 2020] A recommendation stakeholder is any group or individual that can affect, or is affected by, the delivery of recommendations to users 14 Consumers Providers System C P S Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 15. A sample multi-sided scenario 15 Consumers Students Providers Teachers System Online Course Platform Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 16. Multi-sided recommendation aspects [Abdollahpouri et al. 2020] 16 Aspect Definition Multi-stakeholder design A multistakeholder design process is one in which different recommendation stakeholder groups are identified and consulted in the process of system design Multi-stakeholder algorithm A multistakeholder recommendation algorithm takes into account the preferences of multiple parties when generating recommendations, especially when these parties are on different sides of the recommendation interaction Multi-stakeholder evaluation A multistakeholder evaluation is one in which the quality of recommendations is assessed across multiple groups of stakeholders, in addition to a point estimate over the full user population Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Recommendation principles
  • 17. Part II Data and algorithmic bias fundamentals
  • 18. Motivating example in music [Mehrotra et al. 2018] ● People frequently listen to music online ● Ratings and frequencies often used to learn patterns ● 1/3 of users listen to at least 20% of unpopular artists ● Why are popular artists favoured? ● Why do users who tend to interact with niche artists receive the worst recommendations? 18 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals > Motivating examples
  • 19. Motivating example in education [Boratto et al. 2019] ● Online course platforms are receiving great attention ● Student's preferences learnt from ratings/enrolments ● The imbalance in popularity among courses reinforces coverage and concentration biases of ranked courses ● Popularity bias could impede new courses to emerge ● The market could be dominated by a few teachers 19 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals > Motivating examples
  • 20. Motivating example in recruiting [Singh et al. 2018] 20 ● Recruiters rely more and more on automated systems ● Based on the job, "best" candidates are suggested ● Small differences in relevance can lead to large differences in exposure among candidate groups ● Is this winner-take-all allocation of exposure fair, even if the winner just has a tiny advantage in relevance? ● It might be fairer to distribute exposure proportional to relevance, even if this leads to a drop in utility Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals > Motivating examples
  • 21. Disclaimers 21 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● We aim to focus on scientific literature that specifically consider recommender systems ● Pointers to representative scientific events on related concepts applied to ranking systems are given ● References discussed throughout the slides would not be exhaustive ● Refer to the extended bibliography attached to this tutorial for a more comprehensive list of papers ● This tutorial will not cover biases associated to sensitive attributes (unfairness), which have a dedicated ICDE tutorial Data and algorithmic bias fundamentals > Motivating examples
  • 22. Related scientific venues and initiatives 22 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Scientific tutorials including: Dedicated workshops or tracks including: Papers in top-tier conferences including: ● RecSys ● SIGIR ● The Web Conf ● TREC ● UMAP ● WSDM ● CIKM ● ECIR ● KDD ● FAccTRec @ RecSys 2018-2020 ● RMSE & Impact RS @ RecSys 2019 ● FACTS-IR @ SIGIR 2019 ● FATES @ The Web Conf 2019-21 ● FAIR-TREC @ TREC 2020-21 ● FairUMAP @ UMAP 2018-20 ● DAB @ CIKM 2017-19 ● Bias @ ECIR 2020-21 ● FAT-ML @ ICML 2014-2019 ● Fairness and Discrimination in Retrieval and Recommendation @ SIGIR 2019 & RecSys 2019 ● Learning to Rank in theory and practice: From Gradient Boosting to Neural Networks and Unbiased Learning @ SIGIR 2019 ● Multi-stakeholder Recommendations: Case Studies, Methods and Challenges @ RecSys 2019 ● Experimentation with fairness-aware recommendation using librec-auto @ FAT 2020 ● Hands-on on Data and Algorithmic Bias in Recommender Systems @ UMAP 2020 ● Bias in Personalized Rankings: Concepts to Code @ ICDM 2020 Data and algorithmic bias fundamentals Special issues in journals including: ● Special Issue on Fair, Accountable, and Transparent Recommender Systems @ UMUAI ● Special Issue on Algorithmic Bias and Fairness in Search and Recommendation @ IPM
  • 23. Perspectives impacted by bias 23 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Economics Law Rights Security Social Dynamics bias can introduce disparate impacts among providers, influencing future success and revenues bias can affect core user's rights that are regulated by law, such as fairness and discrimination bias can reinforce discrimination of certain user's groups, including ageism, sexism, homophobia bias can lead certain groups of users or an entire system to be more vulnerable to attacks (e.g., bribery) bias can influence how labour markets evolve and can be amplified as the technology progresses Data and algorithmic bias fundamentals > Shaping the context
  • 24. How biases can influence economy [Liu et al. 2019b] ● Loan recommender systems are designed to assist lenders in looking for promising borrowers ● Such systems model lenders’ historical behaviors and generate personalized recommendations ● Certain geographical regions such as Asia and Africa dominate recommendation ● Others like Oceania and Eastern Europe barely receive recommendations 24 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals > Shaping the context
  • 25. Biases that impact on law [Tolan et al. 2019] The right to non-discrimination, which can be undermined by inherent biases, is embedded in the normative framework of the European Union, e.g.: ● Mentions can be found in Art. 21 of the EU Charter of Fundamental Rights ● Article 14 of the European Convention on Human Rights ● Articles 18-25 of the Treaty on the Functioning of the European Union 25 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals > Shaping the context
  • 26. Biases that impact on rights [Yao et al. 2017] As an example, United Nations Sustainable Development Goal 4 aims also to ensure equitable quality education for all, but Yao et al. observed that: ● In 2010, women accounted for only 18% of the BSc degrees in Computer Science ● Historical rating data of Computer Science courses to be dominated by men ● The model may underestimate women’s preferences and be biased towards men 26 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals > Shaping the context
  • 27. ● Sellers might attack the system by introducing a bias in the ratings ● The attack goal is to bribe users to increase ratings and push recommendations ● A novel Hybrid KNN CF is introduced to deal with this phenomenon ● By means of this novel algorithm: ○ The profitability associated to increasing the ratings is strongly reduced w.r.t. SVD ○ Downgrading the ratings of competitors is not profitable with this approach ○ System is more robust to attacks and more trustable by the users Biases can affect security aspects [Ramos et al. 2020] 27 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals > Shaping the context
  • 28. Biases associated to social dynamics [Fabbri et al. 2020] 28 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● People recommendation in social networks, with users divided into groups based on gender ● A range of state-of-the-art algorithms, such as Adamic-Adar, SALSA, and ALS are inspected ● People recommenders produce disparate visibility on the two subgroups of gender ● Homophily plays a key role in promoting or reducing visibility for different subgroups Lorenz Curves (inequality). Recommendations introduce more inequality than degree distribution, and this inequality is stronger in the minority class. Data and algorithmic bias fundamentals > Shaping the context
  • 29. Ethical aspects influenced by bias [Bozdag 2013, Milano et al. 2020] 29 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Moral recommendation of inappropriate content Opacity black-box algorithms, uninformative explanations, feedback effects Privacy unauthorised data collection, data leaks, unauthorised inferences Fairness observation bias, population imbalance Autonomy and Identity behavioural traps and encroachment on sense of personal autonomy Social Exposure lack of exposure to contrasting viewpoints, feedback effects Data and algorithmic bias fundamentals > Shaping the context
  • 30. On biases against immoral content [Pantakar et al. 2019] 30 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● News recommender system generates awareness on biased news, to possibly avoid fake and politically polarized news ● They propose a news clustering and bias score attached to each news. Recommendation of similar, unbiased content ● With a live-user evaluation, the rankings generated by the algorithm match with the ones the users would generate Data and algorithmic bias fundamentals > Shaping the context
  • 31. Bias affecting users' privacy [Resheff et al. 2018] 31 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● User representations may be used to recover private user information such as gender and age, undermining users' privacy ● A privacy-adversarial framework is proposed to eliminate leakage of private information: an adversarial component is appended to the model for each of the demographic variables we want to obfuscate, so that the user representations are optimized to preclude predicting the variables ● Privacy preserving recommendations, minimal overall adverse effect on recommender performance, fairness of results (all knowledge of the attributes is scrubbed from the representations used by the model) Data and algorithmic bias fundamentals > Shaping the context
  • 32. Influence of bias on autonomy [Arnold et al. 2018] 32 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● Text recommender systems that support creative tasks (writing reviews). Do they exhibit unintentional biases in the support that they offer? ● Contextual recommendations are proposed: (1) selection of the three most likely next words, (2) generation of the most likely phrase continuation for each word ● People who get recommended phrasal text entry shortcuts that are skewed positive, write more positive reviews than when presented with negative-skewed shortcuts Data and algorithmic bias fundamentals > Shaping the context
  • 33. How biases can introduce opacity [Eslami et al. 2020] 33 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● Opaque algorithms sometimes make biased or deceptive decisions, many have called for increased transparency ● They conducted an analysis of 242 users’ online discussions about the Yelp review filtering algorithm ● Users defend this algorithm and its opacity depending on their engagement with and gain from the algorithm ● Then, adding transparency into the algorithm changed users’ attitudes towards the algorithm Data and algorithmic bias fundamentals > Shaping the context (a) A filtered review is presented as “recommended” to the user who wrote it; (b) This review, however, presented for other users as a filtered review.
  • 34. Biased effects on social exposure [Papakyriakopoulos et al. 2020] 34 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● A segment of the political discussions on social networks is shaped by users that are over-proportionally active ● By applying geometric topic modeling on German political comments, they demonstrate that hyperactive users have a significant role in the political discourse ● By training recommender systems, they illustrate that models provide very different suggestions to users, when accounting for or ignoring hyperactive behavior Data and algorithmic bias fundamentals > Shaping the context
  • 35. System objectives influenced by bias [Kaminskas et al. 2017, Namatzadeh et al 2018, Singh et al. 2018] 35 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Utility Recommendation Objectives Novelty Diversity Coverage Serendipity the degree to which recommended items are potentially useful and of interest for the user the degree of attention received by (groups of) items or providers the degree to which the list has valuable items not looked for and generate surprise for the user the degree to which the generated recommendations cover the catalog of available items the degree to which the list of retrieved items covers a broad area of the information space the degree to which items are unknown by the user and/or are different from what the user has seen before Visibility & Exposure Data and algorithmic bias fundamentals > Shaping the context
  • 36. Impact on utility (and trade-offs) [Fu et al. 2020] 36 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● Studying recommendation performance according to the level of activity of users: ○ Inactive users are more susceptible to unsatisfactory recommendations (insufficient training data) ○ Recommendations are biased by the training records of more active users ● They proposed an explainable CF + re-ranking to balance predictions and group/individual fairness ● The disparity in utility is reduced while preserving recommendation quality Data and algorithmic bias fundamentals > Shaping the context
  • 37. Impact on coverage (and trade-offs) [Dean et al. 2020] 37 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● The amount of recourse available to a user is the percentage of unseen items that are reachable ● The availability of items in a recommender system is the percentage of items reachable by a user ● Study on linear preference models (SLIM and MF) ● Unavailable items are less popular than available items ● Users with smaller history have more available recourse Data and algorithmic bias fundamentals > Shaping the context
  • 38. Impact on diversity (and trade-offs) [Lee and Hosanagar 2019] ● They investigated the impact of collaborative recommender algorithms commonly used in e-commerce on sales diversity ● The use of traditional collaborative filters is associated with a decrease in sales diversity relative to a world without product recommendations ● The decrease in aggregate sales diversity may not always be accompanied by a corresponding decrease in individual-level consumption diversity 38 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals > Shaping the context
  • 39. Impact on novelty (and trade-offs) [Boratto et al. 2021] 39 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals > Shaping the context ● They show that the probability of being recommended and the item true positive rate are biased against the item popularity ● They propose an in-processing approach aimed at minimizing the biased correlation between user-item relevance and item popularity ● With small losses in accuracy, their popularity-mitigation approach leads to important gains in beyond-accuracy recommendation quality, especially novelty
  • 40. Impact on providers' exposure [Boratto et al. 2020] 40 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● Certain minority groups of providers are being disproportionately affected by unintentional discrimination ● They show how adding observation upsampling can dramatically change the exposure given to providers ● The ranked lists then provide fairer exposure, wider minority-group item coverage, and limited loss in utility Data and algorithmic bias fundamentals > Shaping the context
  • 41. Part III Bias and mitigation through the pipeline
  • 42. Recommendation pipeline 42 Platform Data Model Recommendations Data Preparation Model Prediction Recommendation Delivering Model Evaluation Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Pre- Processed Data Model Setup and Training Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline Data Acquisition and Storage
  • 43. Recommendation pipeline: users 43 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 44. Types of bias associated to users [Olteanu et al. 2017] ● Population biases differences in demographics between a population of users represented in a dataset/platform and a target population ● Behavioral biases differences in user behavior across platforms or contexts, or across users represented in different datasets ● Content biases behavioral biases that are expressed as lexical, syntactic, semantic, and structural differences in the contents generated by users ● Linking biases behavioral biases expressed as differences in the attributes of networks obtained from user connections, interactions or activity ● Temporal biases differences in populations or behaviors over time 44 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 45. Bias on items due to their popularity [Jannach et al. 2014] ● Context: movies, books, hotels, and mobile games ● Algorithms: CB-Filtering, SlopeOne, User-KNN, Item-KNN, FM, RfRec, Funk-SVD, Koren-MF, ALS, BPR ● Findings: techniques performing well on accuracy focus their recommendations on a tiny fraction of the item spectrum or recommend mostly top sellers 45 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 46. On behavioral bias towards popularity [Cañamares and Castells 2018] ● Context: movies ● Algorithms: User KNN, Item KNN, AvgRating, MF, Random, Pop ● Findings: effectiveness or ineffectiveness of popularity depends on the interplay of three main variables: item relevance, item discovery by users, and the decision by users to interact with discovered items. Authors identify the key probabilistic dependencies among these factors 46 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 47. Behavioral patterns on item categories [Guo and Dunson 2015, Lin et al. 2019] ● Items of different genres have different rating values and different samples ● Bayesian multiplicative probit model to uncover category-wise bias in ratings 47 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● User preferences propagate differently in recommendations, according to movie genre and user gender ● SVD++ and BiasedMF dampen the preference bias for movie genres for both men and women ● WRMF is well-calibrated for Sci-Fi/Crime for both men and women but the behavior is inconsistent for Action/Romance Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 48. Biases conveyed by users' content [Piramuthu et al. 2012, Xu et al. 2018, Dai et al. 2018, Vall et al. 2019] ● Sequential bias: the sequence in which reviews are written play an appreciable role in how the next reviews are written ● Opinion bias: given a user–item pair, the opinion bias is defined as the bias between rating and review. The rating matrix is filled with a linear combination of the rating and the review sentiment ● Textual bias inspects how recommenders are influenced by the fact that words may express different meanings in review context 48 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 49. Recommendation pipeline: platform 49 Platform Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 50. Types of bias in platforms [Olteanu et al. 2017] ● Functional biases biases resulting from platform-specific mechanisms or affordances, that is, the possible actions within each system or environment ● Normative biases biases that are a result of written norms or expectations about unwritten norms describing acceptable patterns of behavior on a given platform ● External biases biases resulting from factors outside the platform, including considerations of socioeconomic status, education, social pressure, privacy concerns, interests, language, personality, and culture ● Non-individual accounts interactions on social platforms that are not produced by individuals, but by accounts representing various types of organizations, or by automated agents 50 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 51. Functional bias in feedback loop [Hofmann et al. 2014] ● Context: how a recommender systems evaluation based on implicit feedback relates to rating-based evaluation, and how evaluation outcomes may be affected by bias in user behavior ● Findings: ○ implicit and explicit evaluation agree well when assumptions agree well (e.g., precision@10) ○ match between assumption on user behavior and explicit evaluation matters – if assumptions are violated, the wrong recommender can be preferred 51 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 52. ● Context: characterize the impact of human-system feedback loop in the context of recommender systems, demonstrating the unintended consequences of algorithmic confounding ● Findings: ○ the recommendation feedback loop causes homogenization of user behavior ○ users experience losses in utility due to homogenization effects ○ the feedback loop amplifies the impact of recommender systems on the distribution of item consumption Impact on homogeneity in the loop [Chaney et al. 2018] 52 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 53. Functional aspects influence choices [Adomavicius et al 2013] ● Context: explore how consumer preferences are impacted by predictions of recommender systems ● Findings: ○ the rating presented by a recommender serves as an anchor for the consumer’s preference ○ viewers’ preference ratings can be significantly influenced by the recommendation received ○ the effect is sensitive to the perceived reliability of a recommender system ○ the effect of anchoring is continuous and linear, operating over a range of system perturbations 53 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 54. Missing data biases internal functions [Pradel et al. 2012] ● Context: study two major biases of the selection of items, i.e., some items obtain more ratings than others (popularity) and positive ratings are observed more frequently than negative ratings (positivity) ● Findings: ○ considering missing data as a form of negative feedback during training may improve performances ○ ...but it can be misleading when testing, favoring popularity more than user preferences 54 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 55. Deceptive expectations bias user views [Elsweiler et al. 2017] ● Context: they explore the feasibility of substituting meals that would typically be recommended to users with similar, healthier dishes, investigating how people perceive and select recipes ● Findings: ○ participants are unable to reliably identify which recipe contains most fat due to their answers being biased by lack of information ○ perception of fat content can be influenced by the information available and, in some cases, misleading cues (image or title) can bias a false impression 55 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 56. On decision biases in user's choices [Teppan and Zanker 2015] ● Context: experimental analysis of the impact of different decision biases like decoy or position effects, as well as risk aversion in positive decision frames ● Findings: risk aversion can be observed in all settings, while position and decoy effects only play a role when risk aversion is not too predominant 56 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 57. Data Acquisition and Storage Recommendation pipeline: data 57 Data Data Preparation Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Pre- Processed Data Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 58. Sources of bias in data collection [Olteanu et al. 2017] ● Data acquisition ○ discouraging data collection by third parties ○ programmatic limitation of access to data (e.g., time, amount, size) ○ not all relevant data captured by the platform or opaque and unclear sampling strategies ● Data querying ○ limited expressiveness of APIs regarding information needs ○ different ways of operationalization of information by APIs ○ influence of keywords on datasets in keyword-based queries ● Data filtering ○ removal of outliers that are relevant for the analysis ○ bounding of analysis due to text filtering operations 58 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 59. Sources of bias in data preparation [Olteanu et al. 2017] ● Data cleaning ○ data representation choices and default values ○ normalization procedures (e.g., based on geographical information) ● Data enrichment ○ subjective and noisy labels due to manual annotations ○ errors due to automatic annotation based on statistical or machine learning ● Data aggregation ○ lose of information due to high-level aggregation ○ spurious patterns of association when data is groups based on certain attributes 59 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 60. Recommendation pipeline: model 60 Model Recommendations Model Prediction Recommendation Delivering Model Evaluation Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Model Setup and Training Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 61. Sources of bias in model exploitation [Olteanu et al. 2017] ● Qualitative analysis ○ data representation choices and default values ○ normalization procedures (e.g., based on geographical information) ● Descriptive analysis ○ research often relying on counting entities or influence of bias and confounders on correlation analysis ● Inference analysis ○ definition of target variables, class labels, or data representations ○ effect of the objective function to the inference task ● Observational analysis ○ peer effects due to platform affordances and conventions ○ selection bias and how treatment effects on results generalizability 61 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 62. Sources of bias in model evaluation [Olteanu et al. 2017, Bellogín et al. 2017] ● Evaluation data selection ○ imbalances of data samples due to their popularity ○ sensitivity to the ratio of the test ratings versus the added non-relevant items ● Metrics selection ○ influence of choice of metrics on research study takeaways ○ accounting domain impact throughout performance assessment ● Result assessment and interpretation ○ traces and patterns changing with context ○ going beyond studies evaluated on a single dataset or method 62 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 63. Bias on popularity in evaluation [Bellogín et al. 2017] 63 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● First percentile-based approach: ○ dividing items in m popularity percentiles ○ breaking down the computation of accuracy by such percentiles ○ averaging the m obtained values ● Second uniform test approach: ○ formation of data splits where all items have the same amount of test ratings ○ picking a set T of candidate items and a number g of test ratings per item Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 64. Bias and random decoys in evaluation [Ekstrand et al. 2017] [Other readings: Lim et al. 2015, Yang et al. 2018, Carraro et al. 2020] ● Context: examine the random decoys protocol, where the candidate set consists of the test set items plus a randomly-selected set of N decoy items ● Findings: ○ the distribution of items goodness required to avoid misclassified decoys with reasonable probability is unreasonable ○ there is a serious discrepancy between theoretical and observed behavior of the random decoy strategy with respect to popularity bias 64 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 65. Bias on error estimation in evaluation [Tian et al. 2020] ● Context: offline evaluation cannot accurately assess novel, relevant recommendations; the most novel items are missing from the data and cannot be judged as relevant ● Findings: ○ missing data in the observation process causes the evaluation to mis-estimate metric values ○ substantial breakthroughs in recommendation quality will be difficult to be assessed offline 65 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias through the pipeline
  • 66. Bias-aware process pipeline 66 IDENTIFY PRODUCT GOALS ● What are you trying to achieve? ● For what population of people? ● What metrics are you tacking? MITIGATE ISSUES ● Does data include enough minority samples? ● Do our proxies measure what we think they do? ● Does the bias notion capture stakeholders’ needs? IDENTIFY STAKEHOLDERS ● Who has a stake in this product? ● Who might be harmed? ● How? DEVELOP AND ANALYZE THE SYSTEM ● How well the system matches product goals? ● To what degree bias is still present? ● How decisions impact on each stakeholder? DEFINE A BIAS NOTION ● What type of bias? At what point? ● What distributions? Bias-aware process pipeline Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 67. Techniques for bias treatment 67 Pre-processing before model training In-processing during model training Post-processing after model training Pre-processing techniques try to transform the data so that the bias is mitigated. If the algorithm is allowed to modify the training data, pre-processing can be used In-processing techniques try to modify learning algorithms to mitigate bias during training process. If it is allowed to change the learning procedure, in-processing can be used Post-processing is performed by re-ranking items of the lists obtained after model training. If the algorithm can treat the learned model as a black box, post-processing can be used Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 68. Example of pre-processing treatment [Jannach et al. 2015] ● Idea: extending BPR with a modified distribution function φ that samples tuples (u,i, j) where i is less popular and j is more popular ● Algorithms: BPR ● Findings: it is possible to mitigate popularity bias, with minor losses in effectiveness 68 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 69. Example of in-processing treatment [Abdollahpouri et al. 2017] ● Idea: identify a regularization component of the objective to be minimized when the distribution of recommendations achieves a 50/50 balance between medium-tail and short-head items. ● Algorithms: RankALS (i.e., pair-wise learning) ● Findings: it is possible to model the trade-off between long-tail catalog coverage and ranking accuracy as a multi-objective optimization problem based on a dissimilarity matrix 69 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 70. Example of post-processing treatment [Abdollahpouri et al. 2019] ● Idea: they modified the xQuAD to produce a new re-ranked list S (|S | < |R|) that manages popularity bias while still being accurate ● Algorithms: RankALS (i.e., pair-wise learning) ● Findings: on two datasets, the re-ranking methods boost long-tail items while keeping the accuracy loss small, compared to the model-based technique 70 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 71. Other treatments for popularity bias among others... ● Treatments that manipulate interactions before training a model: ○ sample tuples (u,i, j) where i is less popular than j for pair-wise learning [Jannach et al. 2014] ○ remove popular items, simulating situations in which these items are missing [Cremonesi et al. 2014] ○ detect and fix noisy ratings by characterizing items and users by their profiles [Toledo et al. 2015] ● Treatments that regularize the loss function score during training: ○ a regularization that balance recommendation accuracy and intra-list diversity [Abdollahpouri et al. 2017] ○ a regularization that minimizes the correlation between accuracy and item popularity [Boratto et al. 2020a] ○ adversarial framework: minimax game between the BPR model and a discriminator [Zhu et al. 2020] ● Treatments that re-rank items after model training: ○ two-way aggregation of direct and reversed rank results (to improve coverage and accuracy) [Dong et al. 2020] ○ a re-ranking that suggests first items from unseen providers (to improve coverage) [Burke et al. 2016] ○ a re-ranking score that balances predicted rating with the inverse of popularity [Abdollahpouri et al. 2018] ○ a re-ranking that includes long-tail items the user might like [Abdollahpouri et al. 2019] 71 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 72. Treatments against other biases (1) ● Biases related to how items are sampled, positioned, and/or selected, e.g.: ○ connect recommendation to causal inference from experimental and observational data [Schnabel et al. 2016] ○ integrate imputed errors and propensities, for alleviating the effect of propensity variance [Wang et al. 2019] ○ spiral of silence effect, i.e., users are likely to rate if they perceive a support by the dominant opinion [Liu D. et al. 2019a] ○ estimate item frequency from a data stream, subject to vocabulary and distribution shifts [Yi et al. 2019] ○ model position-bias offline and conduct online inference without position information [Guo et al. 2019] ○ off-policy correction to learn from feedback given by an ensemble of prior model policies [Chen et al. 2019a] ○ a clipped estimator to improve the bias-variance trade-off than w.r.t. unbiased estimator [Saito et al. 2020] ○ a counterfactual approach which accounts for selection and position bias jointly [Ovaisi et al. 2020] ○ a two-stage off-policy that takes the model into account while training the candidate model [Ma et al. 2020] 72 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 73. Treatments against other biases (2) ● Biases associated to how items reach their audience: ○ a novel probabilistic method for weighted sampling of k neighbors that considers the similarity levels between the target user (or item) and the candidate neighbors [Adamopoulos et al. 2014] ○ a target customer re-ranking algorithm to adjust the population distribution and composition in the top-k target customers of an item while maintaining recommendation quality [Zhao et al. 2020] ● Biases associated to how items are marketed, e.g.: ○ a fairness-aware framework to address market imbalance bias by calibrating the parity of prediction errors across different market segments [Wan et al. 2020] 73 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 74. Treatments against other biases (3) ● Biases associated to reviews and textual opinions, e.g.: ○ a sentiment classification scoring method, which employs dual attention vectors to predict the users’ sentiment scores of their reviews , to catch opinion bias and enhance user-item matrix [Xu et al. 2018] ○ a hybrid model that integrates modified-sied information related to textual bias and rating bias in matrix factorization, getting a specific word representation for each item review [Dai et al. 2018] 74 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 75. Treatments against other biases (4) ● Biases associated to social trust and influence: ○ a mitigation using polynomial regression and a Bayesian information criterion to predict ratings less influenced by the tendency to conform to the perceived “norm” in a community [Krishnan et al. 2014] ○ clustering user-item space to discover rating bubbles derived from the theory of social bias, i.e., existing ratings indirectly influences the users' opinion to follow the herd instinct [Divyaa et al. 2019] ○ a matrix completion algorithm that performs hybrid memory-based collaborative filtering, improving how the bribery effect is managed and how the system is robust against bribery [Ramos et al. 2020b] 75 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 76. Treatments against other biases (5) ● Biases related to the interactions of users over time, e.g.: ○ an historical influence-aware latent factor model to capture and mitigate historical distortions in each single rating under the assimilation-contrast theory: users conform to historical ratings if historical ratings are not far from the product quality (assimilation), while users deviate from historical ratings if historical ratings are significantly different from the product quality (contrast) [Zhang et al. 2018] ○ an unbiased loss using inverse propensity weighting, that includes the recency propensity of item x at time t, to be used in point-wise learning to rank [Chen et al. 2019b] 76 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals Data and algorithmic bias fundamentals > Bias Mitigation
  • 78. Steps of this hands on 78 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Hands on recommender systems https://colab.research.google.com/github/biasinrecsys/icde2021/blob/master/notebooks/model_setup.ipynb 1 Data Load We load data from publicly available datasets, specifically focusing on Movielens 1M (movies) Data Pre-Processing We process data to be fed into the model and we prepare training based on point- and pair-wise methods 2 Model Definition and Train We define the architecture of the model, setup the training parameters and run the model training process 3 Relevance Computation Given a pre-trained model, we compute the user-item relevance scores across all the user-item pairs 4 Model Evaluation We compute accuracy and beyond-accuracy metrics, such as coverage, novelty, and diversity 5
  • 79. Part V Hands on item popularity bias
  • 80. Steps of this hands on 80 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Investigation on item popularity bias https://colab.research.google.com/github/biasinrecsys/icde2021/blob/master/notebooks/item_popularity_bias.ipynb 1 Model Exploration We consider on data and models introduced in the first hands on to inspect of popularity impacts on visibility and exposure of items Mitigation Setup We arrange a representative set of mitigation strategies against popularity bias in pre-, in- and post-processing 2 Mitigation Running We run the mitigation procedure, inspecting how the optimization processes influences popularity values 3 Model Re-Evaluation We re-run the evaluation of the first hands on to highlight how disparities among popular and unpopular items are reduced 4 Impact Assessment We interpret the results obtained during evaluation in order to envision how stakeholders are impacted 5
  • 82. Contextual challenges 82 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● Stakeholders have different (and conflicting) needs. How can recommender systems account for them? ● Multi-disciplinary approaches to go beyond algorithms (e.g., to link justice and fairness) ● Synthesizing a definition of bias or fairness is challenging ● Creating a common vocabulary to recognize different types of bias and unfairness ● Data to characterize bias phenomena with enough depth is lacking (especially for sensitive attributes) ● There are forms of bias on the Web that have not been studied in the recommendation literature Concluding Remarks
  • 83. Operational challenges 83 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras ● Measuring and operationalizing bias or fairness. How can we optimize a recommender system for it? ● Can we mitigate multiple forms of bias at the same time? ● Slight changes throughout the pipeline can make a huge difference on impact ● Research and development should be more focused on the real world application ● When mitigating bias we usually trade for other qualities. ● How can we mitigate bias without compromising recommendation quality? Concluding Remarks
  • 84. Evaluation challenges ● What if we do not have the sensitive attributes in the collected data? ● How should we select an approach with respect to another (e.g., equity vs equality)? ● How to identify harms in the considered context? ● Will the chosen offline metrics and experiments lead to the desired results online? ● How to inspect whether data generation and collection methods are appropriate? ● How could we take into account both bias goals and efficiency in the real world? 84 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras Concluding Remarks
  • 85. Resources from this tutorial 1. Tutorial website https://biasinrecsys.github.io/icde2021/ 2. Github repository https://github.com/biasinrecsys/icde2021 3. Jupyter notebooks 3.1. https://colab.research.google.com/github/biasinrecsys/icde2021/blob/master/notebooks/model_setup.ipynb 3.2. https://colab.research.google.com/github/biasinrecsys/icde2021/blob/master/notebooks/item_popularity_bias.ipynb 85 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 87. References #1 1. Himan Abdollahpouri, Gediminas Adomavicius, Robin Burke, Ido Guy, Dietmar Jannach, Toshihiro Kamishima, Jan Krasnodebski, Luiz Augusto Pizzato: Multi Stakeholder recommendation: Survey and research directions. User Model. User Adapt. Interact. 30(1): 127-158 (2020). 2. Himan Abdollahpouri, Robin Burke, Bamshad Mobasher: Managing Popularity Bias in Recommender Systems with Personalized Re-Ranking. FLAIRS Conference 2019: 413-418 (2019). 3. Himan Abdollahpouri, Robin Burke, Bamshad Mobasher. Popularity-Aware Item Weighting for Long-Tail Recommendation. arXiv preprint arXiv:1802.05382 (2018). 4. Himan Abdollahpouri, Robin Burke, Bamshad Mobasher: Controlling Popularity Bias in Learning-to-Rank Recommendation. RecSys 2017: 42-46 (2017). 5. Panagiotis Adamopoulos, Alexander Tuzhilin: On over-specialization and concentration bias of recommendations: probabilistic neighborhood selection in collaborative filtering systems. RecSys 2014: 153-160 (2014). 6. Gediminas Adomavicius, Jesse C. Bockstedt, Shawn P. Curley, Jingjing Zhang: Do Recommender Systems Manipulate Consumer Preferences? A Study of Anchoring Effects. Inf. Syst. Res. 24(4): 956-975 (2013). 7. Vito Walter Anelli, Tommaso Di Noia, Eugenio Di Sciascio, Azzurra Ragone, Joseph Trotta: Local Popularity and Time in top-N Recommendation. ECIR (1) 2019: 861-868 (2019). 8. Kenneth C. Arnold, Krysta Chauncey, Krzysztof Z. Gajos: Sentiment Bias in Predictive Text Recommendations Results in Biased Writing. Graphics Interface 2018: 42-49 (2018). 87 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 88. References #2 9. Alejandro Bellogín, Pablo Castells, Iván Cantador: Statistical biases in Information Retrieval metrics for recommender systems. Inf. Retr. J. 20(6): 606-634 (2017). 10. Alex Beutel, Jilin Chen, Tulsee Doshi, Hai Qian, Li Wei, Yi Wu, Lukasz Heldt, Zhe Zhao, Lichan Hong, Ed H. Chi, Cristos Goodrow: Fairness in Recommendation Ranking through Pairwise Comparisons. KDD 2019: 2212-2220 11. Asia J. Biega, Krishna P. Gummadi, Gerhard Weikum: Equity of Attention: Amortizing Individual Fairness in Rankings. SIGIR 2018: 405-414 12. Black, J. S., & van Esch, P. (2020). AI-enabled recruiting: What is it and how should a manager use it?. Business Horizons, 63(2), 215-226. 13. Ludovico Boratto, Gianni Fenu, Mirko Marras: The Effect of Algorithmic Bias on Recommender Systems for Massive Open Online Courses. ECIR (1) 2019: 457-472 (2019). 14. Ludovico Boratto, Gianni Fenu, Mirko Marras: Connecting User and Item Perspectives in Popularity Debiasing for Collaborative Recommendation. CoRR abs/2006.04275 (2020a). 15. Ludovico Boratto, Gianni Fenu, Mirko Marras: Interplay between Upsampling and Regularization for Provider Fairness in Recommender Systems. CoRR abs/2006.04279 (2020b). 16. Engin Bozdag: Bias in algorithmic filtering and personalization. Ethics Inf Technol 15, 209–227 (2013). 17. Robin Burke. Multisided fairness for recommendation. arXiv preprint arXiv:1707.00093 (2017). 18. Robin Burke, Nasim Sonboli, Aldo Ordonez-Gauger: Balanced Neighborhoods for Multi-sided Fairness in Recommendation. FAT 2018: 202-214 (2018) 19. Robin D. Burke, Himan Abdollahpouri, Bamshad Mobasher, Trinadh Gupta: Towards Multi-Stakeholder Utility Evaluation of Recommender Systems. UMAP Extended Proceedings (2016). 88 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 89. References #3 20. Rocío Cañamares, Pablo Castells: Should I Follow the Crowd?: A Probabilistic Analysis of the Effectiveness of Popularity in Recommender Systems. SIGIR 2018: 415-424 21. Diego Carraro, Derek Bridge: Debiased offline evaluation of recommender systems: a weighted-sampling approach. SAC 2020: 1435-1442 (2020). 22. L. Elisa Celis, Damian Straszak, Nisheeth K. Vishnoi: Ranking with Fairness Constraints. ICALP 2018: 28:1-28:15 (2017). 23. Roberto Centeno, Ramón Hermoso, Maria Fasli: On the inaccuracy of numerical ratings: dealing with biased opinions in social networks. Inf. Syst. Frontiers 17(4): 809-825 (2015). 24. Allison J. B. Chaney, Brandon M. Stewart, Barbara E. Engelhardt: How algorithmic confounding in recommendation systems increases homogeneity and decreases utility. RecSys 2018: 224-232 (2018). 25. Sushma Channamsetty, Michael D. Ekstrand: Recommender Response to Diversity and Popularity Bias in User Profiles. FLAIRS Conference 2017: 657-660 (2017). 26. Minmin Chen, Alex Beutel, Paul Covington, Sagar Jain, Francois Belletti, Ed H. Chi: Top-K Off-Policy Correction for a REINFORCE Recommender System. WSDM 2019: 456-464 (2019a). 27. Ruey-Cheng Chen, Qingyao Ai, Gaya Jayasinghe, W. Bruce Croft: Correcting for Recency Bias in Job Recommendation. CIKM 2019: 2185-2188 (2019b). 28. Paolo Cremonesi, Franca Garzotto, Roberto Pagano, Massimo Quadrana: Recommending without short head. WWW (Companion Volume) 2014: 245-246 (2014). 89 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 90. References #4 29. Jiao Dai, Mingming Li, Songlin Hu, Jizhong Han: A Hybrid Model Based on the Rating Bias and Textual Bias for Recommender Systems. ICONIP (2) 2018: 203-214 (2018). 30. Sarah Dean, Sarah Rich, Benjamin Recht: Recommendations and user agency: the reachability of collaboratively-filtered information. FAT* 2020: 436-445 (2020). 31. Yashar Deldjoo, Vito Walter Anelli, Hamed Zamani, Alejandro Bellogín, Tommaso DiNoia - A Flexible Framework for Evaluating User and Item Fairness in Recommender Systems. In User Modeling and User-Adapted Interaction (2020) 32. Fernando Diaz, Bhaskar Mitra, Michael D. Ekstrand, Asia J. Biega, Ben Carterette: Evaluating Stochastic Rankings with Expected Exposure. CoRR abs/2004.13157 (2020). 33. Divyaa L. R., Nargis Pervin: Towards generating scalable personalized recommendations: Integrating social trust, social bias, and geo-spatial clustering. Decis. Support Syst. 122 (2019). 34. Qiang Dong, Quan Yuan, Yang-Bo Shi: Alleviating the recommendation bias via rank aggregation. Physica A: Statistical Mechanics and its Applications, 534, 122073. (2019). 35. Bora Edizel, Francesco Bonchi, Sara Hajian, André Panisson, Tamir Tassa: FaiRecSys: mitigating algorithmic bias in recommender systems. Int. J. Data Sci. Anal. 9(2): 197-213 (2020) 36. David Elsweiler, Christoph Trattner, Morgan Harvey: Exploiting Food Choice Biases for Healthier Recipe Recommendation. SIGIR 2017: 575-584 (2017). 37. Michael D. Ekstrand, Mucun Tian, Ion Madrazo Azpiazu, Jennifer D. Ekstrand, Oghenemaro Anuyah, David McNeill, Maria Soledad Pera: All The Cool Kids, How Do They Fit In?: Popularity and Demographic Biases in Recommender Evaluation and Effectiveness. FAT 2018: 172-186 (2018a). 90 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 91. References #5 38. Michael D. Ekstrand, Mucun Tian, Mohammed R. Imran Kazi, Hoda Mehrpouyan, Daniel Kluver: Exploring author gender in book rating and recommendation. RecSys 2018: 242-250 (2018b) 39. Michael D. Ekstrand, Vaibhav Mahant: Sturgeon and the Cool Kids: Problems with Random Decoys for Top-N Recommender Evaluation. FLAIRS Conference 2017: 639-644 (2017). 40. Eslami, M., Vaccaro, K., Lee, M. K., Elazari Bar On, A., Gilbert, E., & Karahalios, K. (2019, May). User attitudes towards algorithmic opacity and transparency in online reviewing platforms. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (pp. 1-14). 41. Francesco Fabbri, Francesco Bonchi, Ludovico Boratto, Carlos Castillo: The Effect of Homophily on Disparate Visibility of Minorities in People Recommender Systems. ICWSM 2020: 165-175 (2020). 42. Golnoosh Farnadi, Pigi Kouki, Spencer Thompson, Sriram Srinivasan, Lise Getoor: A Fairness-aware Hybrid Recommender System. CoRR abs/1809.09030 (2018). 43. Zuohui Fu, Yikun Xian, Ruoyuan Gao, Jieyu Zhao, Qiaoying Huang, Yingqiang Ge, Shuyuan Xu, Shijie Geng, Chirag Shah, Yongfeng Zhang, Gerard de Melo: Fairness-Aware Explainable Recommendation over Knowledge Graphs. CoRR abs/2006.02046 (2020). 44. Huifeng Guo, Jinkai Yu, Qing Liu, Ruiming Tang, Yuzhou Zhang: PAL: a position-bias aware learning framework for CTR prediction in live recommender systems. RecSys 2019: 452-456 (2019). 45. Fangjian Guo, David B. Dunson: Uncovering Systematic Bias in Ratings across Categories: a Bayesian Approach. RecSys 2015: 317-320 (2015). 46. Katja Hofmann, Anne Schuth, Alejandro Bellogín, Maarten de Rijke: Effects of Position Bias on Click-Based Recommender Evaluation. ECIR 2014: 624-630 (2014). 47. Longke Hu, Aixin Sun, Yong Liu: Your neighbors affect your ratings: on geographical neighborhood influence to rating prediction. SIGIR 2014: 345-354 (2014). 91 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 92. References #6 48. Farnaz Jahanbakhsh, Justin Cranshaw, Scott Counts, Walter S. Lasecki, Kori Inkpen: An Experimental Study of Bias in Platform Worker Ratings: The Role of Performance Quality and Gender. CHI 2020: 1-13 (2020). 49. Dietmar Jannach, Lukas Lerche, Iman Kamehkhosh, Michael Jugovac: What recommenders recommend: an analysis of recommendation biases and possible countermeasures. User Model. User Adapt. Interact. 25(5): 427-491 (2015). 50. Marius Kaminskas, Derek Bridge: Diversity, Serendipity, Novelty, and Coverage: A Survey and Empirical Analysis of Beyond-Accuracy Objectives in Recommender Systems. ACM Trans. Interact. Intell. Syst. 7(1): 2:1-2:42 (2017). 51. Toshihiro Kamishima, Shotaro Akaho, Hideki Asoh, Jun Sakuma: Recommendation Independence. FAT 2018: 187-201 (2018). 52. Mesut Kaya, Derek Bridge, and Nava Tintarev. 2020. Ensuring Fairness in Group Recommendations by Rank-Sensitive Balancing of Relevance. In Fourteenth ACM Conference on Recommender Systems (RecSys '20). 53. Dominik Kowald, Markus Schedl, Elisabeth Lex: The Unfairness of Popularity Bias in Music Recommendation: A Reproducibility Study. ECIR (2) 2020: 35-42 (2020). 54. Sanjay Krishnan, Jay Patel, Michael J. Franklin, Ken Goldberg: A methodology for learning, analyzing, and mitigating social influence bias in recommender systems. RecSys 2014: 137-144 (2014). 55. Caitlin Kuhlman, MaryAnn Van Valkenburg, Elke A. Rundensteiner: FARE: Diagnostics for Fair Ranking using Pairwise Error Metrics. WWW 2019: 2936-2942 (2019). 56. Preethi Lahoti, Krishna P. Gummadi, Gerhard Weikum: iFair: Learning Individually Fair Data Representations for Algorithmic Decision Making. ICDE 2019: 1334-1345 92 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 93. References #7 57. Dokyun Lee, Kartik Hosanagar: How Do Recommender Systems Affect Sales Diversity? A Cross-Category Investigation via Randomized Field Experiment. Inf. Syst. Res. 30(1): 239-259 (2019). 58. Jurek Leonhardt, Avishek Anand, Megha Khosla: User Fairness in Recommender Systems. WWW (Companion Volume) 2018: 101-102 (2018). 59. Mingming Li, Jiao Dai, Fuqing Zhu, Liangjun Zang, Songlin Hu, Jizhong Han: A Fuzzy Set Based Approach for Rating Bias. AAAI 2019: 9969-9970 (2019). 60. Daryl Lim, Julian J. McAuley, Gert R. G. Lanckriet: Top-N Recommendation with Missing Implicit Feedback. RecSys 2015: 309-312 (2015). 61. Xiao Lin, Min Zhang, Yongfeng Zhang, Zhaoquan Gu, Yiqun Liu, Shaoping Ma: Fairness-Aware Group Recommendation with Pareto-Efficiency. RecSys 2017: 107-115 (2017). 62. Kun Lin, Nasim Sonboli, Bamshad Mobasher, Robin Burke: Crank up the Volume: Preference Bias Amplification in Collaborative Recommendation. RMSE@RecSys 2019 (2019). 63. Dugang Liu, Chen Lin, Zhilin Zhang, Yanghua Xiao, Hanghang Tong: Spiral of Silence in Recommender Systems. WSDM 2019: 222-230 (2019a). 64. Weiwen Liu, Jun Guo, Nasim Sonboli, Robin Burke, Shengyu Zhang: Personalized fairness-aware re-ranking for microlending. RecSys 2019: 467-471 (2019b). 65. Jiaqi Ma, Zhe Zhao, Xinyang Yi, Ji Yang, Minmin Chen, Jiaxi Tang, Lichan Hong, Ed H. Chi: Off-policy Learning in Two-stage Recommender Systems. WWW 2020: 463-473 (2020). 66. Benjamin M. Marlin, Richard S. Zemel, Sam T. Roweis, Malcolm Slaney: Recommender Systems, Missing Data and Statistical Model Estimation. IJCAI 2011: 2686-2691 (2011). 93 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 94. References #8 68. Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, Aram Galstyan. A survey on bias and fairness in machine learning. arXiv preprint arXiv:1908.09635 (2019). 69. Rishabh Mehrotra, James McInerney, Hugues Bouchard, Mounia Lalmas, Fernando Diaz: Towards a Fair Marketplace: Counterfactual Evaluation of the trade-off between Relevance, Fairness & Satisfaction in Recommendation Systems. CIKM 2018: 2243-2251 (2018). 70. Silvia Milano, Mariarosaria Taddeo, Luciano Floridi. Recommender systems and their ethical challenges. AI & Soc (2020). 71. Natwar Modani, Deepali Jain, Ujjawal Soni, Gaurav Kumar Gupta, Palak Agarwal: Fairness Aware Recommendations on Behance. PAKDD (2) 2017: 144-155 (2017). 72. Azadeh Nematzadeh, Giovanni Luca Ciampaglia, Filippo Menczer, Alessandro Flammini: How algorithmic popularity bias hinders or promotes quality. CoRR abs/1707.00574 (2017). 73. Alexandra Olteanu, Carlos Castillo, Fernando Diaz, Emre Kiciman: Social Data: Biases, Methodological Pitfalls, and Ethical Boundaries. Frontiers Big Data 2: 13 (2019) 74. Zohreh Ovaisi, Ragib Ahsan, Yifan Zhang, Kathryn Vasilaky, Elena Zheleva: Correcting for Selection Bias in Learning-to-rank Systems. WWW 2020: 1863-1873 (2020). 75. Papakyriakopoulos, Orestis, Juan Carlos Medina Serrano, and Simon Hegelich. "Political communication on social media: A tale of hyperactive users and bias in recommender systems." Online Social Networks and Media 15 (2020): 100058. 76. Anish Anil Patankar, Joy Bose, Harshit Khanna: A Bias Aware News Recommendation System. ICSC 2019: 232-238 (2019). 77. Bruno Pradel, Nicolas Usunier, Patrick Gallinari: Ranking with non-random missing ratings: influence of popularity and positivity on evaluation metrics. RecSys 2012: 147-154 (2012). 94 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 95. References #9 78. Gourab K. Patro, Arpita Biswas, Niloy Ganguly, Krishna P. Gummadi, Abhijnan Chakraborty: FairRec: Two-Sided Fairness for Personalized Recommendations in Two-Sided Platforms. WWW 2020: 1194-1204 (2020). 79. Gourab K. Patro, Abhijnan Chakraborty, Niloy Ganguly, Krishna Gummadi. Incremental Fairness in Two-Sided Market Platforms: On Smoothly Updating Recommendations. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 34, No. 01, pp. 181-188). 80. Selwyn Piramuthu, Gaurav Kapoor, Wei Zhou, Sjouke Mauw: Input online review data and related bias in recommender systems. Decis. Sup. Sys. 53(3): 418-424 (2012). 81. Raghavan, M., Barocas, S., Kleinberg, J., & Levy, K. (2020, January). Mitigating bias in algorithmic hiring: Evaluating claims and practices. In Proceedings of the 2020 conference on fairness, accountability, and transparency (pp. 469-481). 82. Guilherme Ramos and Ludovico Boratto, Reputation (In)dependence in Ranking Systems: Demographics Influence Over Output Disparities, in Proceedings of the 43rd International ACM SIGIR Conference on Researchand Development in Information Retrieval, SIGIR 2020 (2020a) 83. Guilherme Ramos, Ludovico Boratto, Carlos Caleiro. On the negative impact of social influence in recommender systems: A study of bribery in collaborative hybrid algorithms. Information Processing & Management, 57(2), 102058 (2020). 84. Bashir Rastegarpanah, Krishna P. Gummadi, Mark Crovella: Fighting Fire with Fire: Using Antidote Data to Improve Polarization and Fairness of Recommender Systems. WSDM 2019: 231-239 85. Yehezkel S. Resheff, Yanai Elazar, Moni Shahar, Oren Sar Shalom: Privacy and Fairness in Recommender Systems via Adversarial Training of User Representations. ICPRAM 2019: 476-482 86. Francesco Ricci, Lior Rokach, Bracha Shapira: Recommender Systems: Introduction and Challenges. Recommender Systems Handbook 2015: 1-34 (2015). 95 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 96. References #10 87. Yuta Saito, Suguru Yaginuma, Yuta Nishino, Hayato Sakata, Kazuhide Nakata: Unbiased Recommender Learning from Missing-Not-At-Random Implicit Feedback. WSDM 2020: 501-509 (2020). 88. Tobias Schnabel, Adith Swaminathan, Ashudeep Singh, Navin Chandak, Thorsten Joachims: Recommendations as Treatments: Debiasing Learning and Evaluation. ICML 2016: 1670-1679 (2016). 89. Christina Schwind, Jürgen Buder: Reducing confirmation bias and evaluation bias: When are preference-inconsistent recommendations effective - and when not? Comput. Hum. Behav. 28(6): 2280-2290 (2012). 90. Dimitris Serbos, Shuyao Qi, Nikos Mamoulis, Evaggelia Pitoura, Panayiotis Tsaparas: Fairness in Package-to-Group Recommendations. WWW 2017: 371-379 (2017). 91. Shakespeare, D., Porcaro, L., Gómez, E., & Castillo, C. Exploring Artist Gender Bias in Music Recommendation. 92. Ashudeep Singh, Thorsten Joachims: Policy Learning for Fairness in Ranking. NeurIPS 2019: 5427-5437 (2019). 93. Ashudeep Singh, Thorsten Joachims: Fairness of Exposure in Rankings. KDD 2018: 2219-2228 (2018). 94. Nasim Sonboli, Robin Burke: Localized Fairness in Recommender Systems. UMAP (Adjunct Publication) 2019: 295-300 (2019). 95. Harald Steck: Item popularity and recommendation accuracy. RecSys 2011: 125-132 (2011). 96. Maria Stratigi, Haridimos Kondylakis, Kostas Stefanidis: Fairness in Group Recommendations in the Health Domain. ICDE 2017: 1481-1488 (2017). 97. Erich Teppan, Marcus Zanker, M: Decision biases in recommender systems. Journal of Internet Commerce, 14(2), 255-275 (2015). 98. Mucun Tian, Michael D. Ekstrand: Estimating Error and Bias in Offline Evaluation Results. CHIIR 2020: 392-396 (2020). 96 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 97. References #11 99. Songül Tolan: Fair and Unbiased Algorithmic Decision Making: Current State and Future Challenges. CoRR abs/1901.04730 (2019). 100. Raciel Yera Toledo, Yailé Caballero Mota, Luis Martínez-López: Correcting noisy ratings in collaborative recommender systems. Knowl. Based Syst. 76: 96-108 (2015). 101. Andreu Vall, Massimo Quadrana, Markus Schedl, Gerhard Widmer: Order, context and popularity bias in next-song recommendations. Int. J. Multim. Inf. Retr. 8(2): 101-113 (2019). 102. Mengting Wan, Jianmo Ni, Rishabh Misra, Julian J. McAuley: Addressing Marketing Bias in Product Recommendations. WSDM 2020: 618-626 (2020). 103. Xiaojie Wang, Rui Zhang, Yu Sun, Jianzhong Qi: Doubly Robust Joint Learning for Recommendation on Data Missing Not at Random. ICML 2019: 6638-6647 (2019). 104. Jacek Wasilewski, Neil Hurley: Are You Reaching Your Audience?: Exploring Item Exposure over Consumer Segments in Recommender Systems. UMAP 2018: 213-217 (2018). 105. Bin Xia, Junjie Yin, Jian Xu, Yun Li: WE-Rec: A fairness-aware reciprocal recommendation based on Walrasian equilibrium. Knowl. Based Syst. 182 (2019). 106. Yuanbo Xu, Yongjian Yang, Jiayu Han, En Wang, Fuzhen Zhuang, Hui Xiong: Exploiting the Sentimental Bias between Ratings and Reviews for Enhancing Recommendation. ICDM 2018: 1356-1361 (2018). 107. Himank Yadav, Zhengxiao Du, Thorsten Joachims: Fair Learning-to-Rank from Implicit Feedback. CoRR abs/1911.08054 (2019). 108. Longqi Yang, Yin Cui, Yuan Xuan, Chenyang Wang, Serge J. Belongie, Deborah Estrin: Unbiased offline recommender evaluation for missing-not-at-random implicit feedback. RecSys 2018: 279-287 (2018). 97 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras
  • 98. References #12 110. Ke Yang, Julia Stoyanovich: Measuring Fairness in Ranked Outputs. SSDBM 2017: 22:1-22:6 (2017). 111. Sirui Yao, Bert Huang: Beyond Parity: Fairness Objectives for Collaborative Filtering. NIPS 2017: 2921-2930 (2017). 112. Xinyang Yi, Ji Yang, Lichan Hong, Derek Zhiyuan Cheng, Lukasz Heldt, Aditee Kumthekar, Zhe Zhao, Li Wei, Ed H. Chi: Sampling-bias-corrected neural modeling for large corpus item recommendations. RecSys 2019: 269-277 (2019). 113. Meike Zehlike, Carlos Castillo: Reducing Disparate Exposure in Ranking: A Learning To Rank Approach. WWW 2020: 2849-2855 (2020). 114. Meike Zehlike, Francesco Bonchi, Carlos Castillo, Sara Hajian, Mohamed Megahed, Ricardo Baeza-Yates: FA*IR: A Fair Top-k Ranking Algorithm. CIKM 2017: 1569-1578 (2017). 115. Shuai Zhang, Lina Yao, Aixin Sun, Yi Tay: Deep Learning Based Recommender System: A Survey and New Perspectives. ACM Comput. Surv. 52(1): 5:1-5:38 (2019) 116. Xiaoying Zhang, Hong Xie, Junzhou Zhao, John C. S. Lui: Modeling the Assimilation-Contrast Effects in Online Product Rating Systems: Debiasing and Recommendations. IJCAI 2018: 5409-5413 (2018). 117. Xing Zhao, Ziwei Zhu, Majid Alfifi, James Caverlee: Addressing the Target Customer Distortion Problem in Recommender Systems. WWW 2020: 2969-2975 (2020). 118. Yong Zheng, Tanaya Dave, Neha Mishra, Harshit Kumar: Fairness In Reciprocal Recommendations: A Speed-Dating Study. UMAP 2018: 29-34 (2018). 119. Ziwei Zhu, Xia Hu, James Caverlee: Fairness-Aware Tensor-Based Recommendation. CIKM 2018: 1153-1162 (2018). 120. Ziwei Zhu, Jianling Wang, and James Caverlee. 2020. Measuring and Mitigating Item Under-Recommendation Bias in Personalized Ranking Systems. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '20). 98 Countering Bias in Personalized Rankings: From Data Engineering to Algorithm Development Boratto and Marras