Automating Google Workspace (GWS) & more with Apps Script
Rational and irrational bias in recommendation
1. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
Pablo Castells
Universidad Autónoma de Madrid
http://ir.ii.uam.es/castells
RMIT, June 3, 2021
2. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Outline
1. Fairness and bias
2. Bias in recommendation
3. Majority bias in offline evaluation
4. Conclusion
3. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
1. Fairness and bias
2. Bias in recommendation
3. Majority bias in offline evaluation
4. Conclusion
Outline
4. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Fairness angles
Fairness is complex
Has multiple definitions
Different angles
No universal formula
Fair ≠ uniform
Olteanu et al. Frontiers in Big Data 2019, Baeza Comm. ACM 2018, Ntoutsi et al. Rev. Data Min. Knowl. Discov. 2020
5. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Fairness angles
Equal representation
𝑝 group 𝐴 hire = 𝑝 group 𝐵 hire
Statistical parity
𝑝 hire group 𝐴 = 𝑝 hire group 𝐵
Equal opportunity
𝑝 hire qualifications = 𝑥, group 𝐴 = 𝑝 hire qualifications = 𝑥, group 𝐵
Fair ≠ uniform
Olteanu et al. Frontiers in Big Data 2019, Baeza Comm. ACM 2018, Ntoutsi et al. Rev. Data Min. Knowl. Discov. 2020, Pleiss et al NIPS 2017.
6. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
>
Bias for…
• Rational thinking
• Scientific method
• Axioms of geometry
• Trust on scientific authority
Bias is… bad?
7. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Shared
principles
• Rationalism, ethics, human
rights, well being …
“Truth” is a construct
• Based on method
• Subjective, consensus
• Model-dependent
• Useful
Kant
Kuhn
Hawking
>
Bias for…
• Rational thinking
• Scientific method
• Axioms of geometry
• Trust on scientific authority
Bias is… bad?
Fair ≠ neutral
Our
“axioms”?
8. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Truth vs. truth system
Euclidean
geometry
Hyperbolic
geometry
Elliptic
geometry
Axioms involve a choice
What’s a “rectangle”?
9. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Truth vs. truth system
Let 𝑅 = 𝐴 𝐴 ∉ 𝐴 }
Then 𝑅 ∈ 𝑅 ⇔ 𝑅 ∉ 𝑅 !!
Axioms are not perfect – they are just meant
to be useful
Even in math, multiple alternative axiom
systems exist, and they are often revised
10. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Bias and fairness
My point
• Fairness and bias are arbitrarily complex questions
• The importance of bias awareness and understanding
to better cope with bias
11. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Outline
1. Fairness and bias
2. Bias in recommendation
3. Majority bias in offline evaluation
4. Conclusion
12. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
1. Fairness and bias
2. Bias in recommendation
3. Majority bias in offline evaluation
4. Conclusion
Outline
13. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Recommender systems
? ?
? ? ?
? ?
? ?
14. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Majority bias in recommendation
Best sellers
Long-tail items
Items
Users
Popular
items
Rest of items
(long tail)
Items
#
Interactions
Bias in the data
15. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Majority bias in recommendation
Bias in offline evaluation
Ranking by
# positive ratings
0.3
0.2
0.1
0
nDCG@10
0
0.1
0.2
0.3
Random
Avg.
rating
Nr.
ratings
User-based
Matrix
fact.
nDCG@10
Netflix
Cañamares & Castells SIGIR 2018, Cremonesi et al. RecSys 2010, Jannach et al. UMUAI 2015
Popular items
(short head)
Rest of items
(long tail)
Items
Users
User-item
interaction
Unobserved
preference
16. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Majority bias in recommendation
Cañamares & Castells SIGIR 2018, Cremonesi et al. RecSys 2010, Jannach et al. UMUAI 2015
Popular items
(short head)
Rest of items
(long tail)
Items
Users
Test data
(relevant)
Unobserved
preference
User-item
interaction
P@𝑘 ∼
𝑘 Users
𝑘
Bias in offline evaluation
Ranking by
# positive ratings
0.3
0.2
0.1
0
nDCG@10
0
0.1
0.2
0.3
Random
Avg.
rating
Nr.
ratings
User-based
Matrix
fact.
nDCG@10
Netflix
17. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Majority bias in recommendation
Bias in algorithms
Cañamares & Castells SIGIR 2017, Jannach et al. UMUAI 2015, Zhu et al. WSDM 2021
MovieLens 1M dataset
0
400
800
0 1000 2000
Matrix factorization
# positive ratings
Nº
times
top
10
800
400
0
0 1000 2000
0
1000
2000
0 1000 2000
User-based kNN
# positive ratings
2000
1000
0
0 1000 2000
0
1000
2000
3000
0 1000 2000
Item-based kNN
# positive ratings
3000
1000
0
0 1000 2000
2000
18. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Majority bias in recommendation
• “Bias = distortion”
• Majority → herd behavior
Are we getting it all wrong?
19. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Is the majority bias useful?
Irrational
20. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Is the majority bias useful?
Should I follow
the crowd? Rational
21. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Popular recommendations
Things to do in Paris
1. Musée d’Orsay
2. Notre-Dame
de Paris
3. Sainte Chapelle
4. Palais Garnier
4. Eiffel Tower
6. Musée l’Orangerie
7. Arc de Triomphe
8. Louvre Museum
9. Montmartre
22. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Popular recommendations
Things to do in Paris
1. Musée d’Orsay
2. Notre-Dame
de Paris
3. Sainte Chapelle
4. Palais Garnier
4. Eiffel Tower
6. Musée l’Orangerie
7. Arc de Triomphe
8. Louvre Museum
9. Montmartre
23. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
1 2 3 4 5 6 7 8 9
10
0
20
30
Songs
0%
5%
10%
How reliable is the majority bias? Salganik, Dodd & Watt’s experiment
Salganik et al. Science 2006
➢ Degree of randomness in success of song
➢ Social influence…
• reinforces majority skew
• increases irrationality
48 songs
Listen & freely
download
See download count
Don’t see
download count
14k people in 8+1 parallel “worlds”
Downloads
10%
5%
0%
Worlds
#
downloads
rank
One example
song
24. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Is the majority bias useful? Menczer’s simulation: popularity raking vs. quality
“Follow the crowd” behavior
Avg quality of choice
“Laziness”
(top-heaviness
in
popularity
ranking)
Ciampaglia et al. Sci. Rep. 2018
Simulation:
• People choose items and interact
• People see interaction count
• Parameterized choice behavior:
herd vs. quality, browsing effort
Rational herd area: optimum = mild herd
behavior + moderate effort budget
25. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Is the majority bias reliable
Recommender systems: rational or irrational herd?
Rec.
system
Rec.
system
26. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Outline
1. Fairness and bias
2. Bias in recommendation
3. Majority bias in offline evaluation
4. Conclusion
27. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
1. Fairness and bias
2. Bias in recommendation
3. Majority bias in offline evaluation
4. Conclusion
Outline
28. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
𝑢
Where do biases come from?
Cañamares & Castells SIGIR 2018
Discover Engage
𝑖
Popularity
distribution
𝑝 𝑂 𝑖
Recommender
system
Recommend
Your ad here
Advertisement
Friends
Random
chance
Search
Observe
External
factors
The feedback loop
Internal
factors
Relevant
29. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Internal
factors
Relevant
Where do biases come from?
Cañamares & Castells SIGIR 2018
Popularity
distribution
𝑝 𝑂 𝑖
Recommender
system
Your ad here
Advertisement
Friends
Random
chance
Search
External
factors
𝑢
Engage
𝑖
Recommend
Observe
Discover
Retrain
(LTR)
The feedback loop
30. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
The feedback loop
Impressions
High rank
User
engagement
Model
training
More impressions
Higher rank
More user
engagement
Model
training
· · ·
Model
training
Offline evaluation
improvement
Algorithm selection
re-confirmed
Recommend
Recommend
31. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Remove the bias
In the metrics
Stratified recall
Off-policy evaluation
In the algorithms
Steck RecSys 2011, Schnabel et al. ICML 2016, Swaminathan et. al NIPS 2017, Yang et al. RecSys 2018, etc.
Test data subsampling procedures
Unbiased datasets, e.g. Yahoo! R3, Coat, CM100K
In the data
Counterfactual learning
Multi-armed bandits
32. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Remove the bias
In the metrics
Stratified recall
Off-policy evaluation
In the algorithms
Steck RecSys 2011, Schnabel et al. ICML 2016, Swaminathan et. al NIPS 2017, Yang et al. RecSys 2018, etc.
Test data subsampling procedures
Unbiased datasets, e.g. Yahoo! R3, Coat, CM100K
In the data
Counterfactual learning
Multi-armed bandits
• Model/learn the
observation bias 𝑝 𝑂𝑢.𝑖
• Factor it out
: explore beyond the feedback loop
33. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Biased vs. unbiased evaluation
10 random tracks per user
1,000 music tracks
Missing data
5,400
users
Yahoo! R3 dataset
Free user interaction
130K ratings
Marlin & Zemel RecSys 2009
Non-random data
Random data
34. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Biased vs. unbiased evaluation
Cañamares & Castells SIGIR 2018, Mena et al. ACM TOIS in press
Unbiased evaluation Biased evaluation
Users
Non-random
data
Random data
Items
Training
Missing data
Test
Items
Training
Missing data
Users
35. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Biased vs. unbiased evaluation
Mena et al. ACM TOIS in press, Cañamares & Castells SIGIR 2018
Yahoo! R3 – 16 algorithms
Biased P@10
Unbiased
P@10
Pop
Item KNN
LRMF
Rnd
SLIM
SVD++
RSGD
User KNN
Avg
PNMF
EALS
GBPR
GPLSA
BPMF
WRMF
WBPR
0
0.002
0.004
0.006
0.008
0.01
0 0.02 0.04 0.06
Real
P
@10
Where is the bias?
• Why such agreement?
• Can we always expect this?
• When and when not?
36. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Conditional (in)dependences between variables
Cañamares & Castells SIGIR 2018
𝑢
Engage Observe
Relevant
Popularity
distribution
𝑝 𝑂 𝑖
0.3
0.2
0.1
0
nDCG@10
0
0.1
0.2
0.3
Random
Avg.
rating
Nr.
ratings
User-based
Matrix
fact.
nDCG@10
Netflix
Evaluation
Recommender
system
Discover
𝑖
𝑅
𝑂
𝑖
𝑅
𝑂
37. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
𝑢
Conditional (in)dependences between variables
1. Observation depends
just on relevance
2. Observation independent
from relevance
3. Observation depends
both on the item and relevance
Discover
Popularity
distribution
𝑝 𝑂 𝑖
0.3
0.2
0.1
0
nDCG@10
0
0.1
0.2
0.3
Random
Avg.
rating
Nr.
ratings
User-based
Matrix
fact.
nDCG@10
Netflix
Evaluation
Engage
Recommender
system
𝑖
𝑅
𝑂 𝑖
𝑅
𝑂
𝑖
𝑅
𝑂
Relevant
Observe
𝑝 𝑂 𝑅, 𝑖 = 𝑝 𝑂 𝑅 𝑝 𝑂 𝑅, 𝑖 = 𝑝 𝑂 𝑖
𝑖
𝑅
𝑂
38. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Harmless bias
Cañamares & Castells SIGIR 2018
Unbiased
P@10
Biased P@10
Comparisons are preserved
Rational majority emerges
1. Observation depends
strongly on relevance
(simulation)
• You find items through search engines,
good recommender systems, good friends
• You engage based on whether or not
you like the choices
𝑖 𝑅 𝑂
39. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
• Heavy advertisement
• Conformity, manipulation
• Reinforcement loops
Harmless vs. misleading bias
0
0.1
0.2
0.3
0.4
0 0.1 0.2 0.3 0.4 0.5
𝑝 𝑂 𝑖
Uncorrelated
observation
and relevance
𝑝
𝑅
𝑖
Unbiased
P@10
Biased P@10
Just noise
Cañamares & Castells SIGIR 2018, Mena et al. ACM TOIS in press
0
0.1
0.2
0.3
0.4
0 0.1 0.2 0.3 0.4 0.5
𝑝 𝑂 𝑖
𝑝
𝑅
𝑖
Observation
opposes
relevance
Unbiased
P@10
Biased P@10
2. Observation independent
from relevance given item
(simulations)
𝑖
𝑅
𝑂
Misleading bias
40. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Misleading bias
• Negativity + virality
• Item quality increase over time
Cañamares & Castells SIGIR 2018, Mena et al. ACM TOIS in press
Misleading crowd effects emerge
Misleading bias
Unbiased
P@10
Biased P@10
0
0.1
0.2
0.3
0.4
0 0.1 0.2 0.3 0.4 0.5
𝑝 𝑂 𝑖
𝑝
𝑅
𝑖
Observation
opposes
relevance
• Heavy advertisement
• Conformity, manipulation
• Reinforcement loops
2. Observation independent
from relevance given item
𝑖
𝑅
𝑂
41. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
General case: real data
Mena et al. ACM TOIS in press
Biased P@10
Unbiased
P@10
Pop
Item KNN
LRMF
Rnd
SLIM
SVD++
RSGD
User KNN
Avg
PNMF
EALS
GBPR
GPLSA
BPMF
WRMF
WBPR
0
0.002
0.004
0.006
0.008
0.01
0 0.02 0.04 0.06
Real
P
@10
Results would suggest that the typical case is a mix of
1. noticeable relevance dependence
2. item dependence with positive relevance correlation
Rational trend
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
𝑝 𝑂 𝑖
𝑝
𝑂
𝑅,
𝑖
Full dependence
on relevance
Yahoo! R3
0
0.1
0.2
0.3
0.4
0 0.1 0.2 0.3 0.4 0.5
𝑝 𝑅 𝑖
𝑝
𝑂
𝑖
𝑖
𝑅
𝑂
3. Observation depends on both relevance and item
42. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Outline
1. Fairness and bias
2. Bias in recommendation
3. Majority bias in offline evaluation
4. Conclusion
43. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
1. Fairness and bias
2. Bias in recommendation
3. Majority bias in offline evaluation
4. Conclusion
Outline
44. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Conclusion and final thoughts
Majority-biased recommendation is rational if…
– user findings and choices are guided by relevance (regardless of positivity/negativity)
– when relevance is ignored, majority rank doesn’t oppose relevance
A mix of those two trends seems common
– The majority bias is useful for recommendation, (broadly) harmless to evaluation
But the majority bias can become misleading when popularity attracts
negativity
45. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Beyond relevance
Bias is not just a relevance issue
Novelty
Fairness
Generality
Equal opportunity: 𝑝 recommend 𝑅, 𝑖 = 𝑝 recommend 𝑅
Affirmative action: relevance builds on exposure
Lost opportunities: new choices, new algorithms,
better choice among relevants
Leave the crowd, take risk and discover
What do we mean by “relevant”, e.g. value/utility vs. rush of adrenaline
Replace “relevance” by “well-being”, “truth”…?
User-specific bias?
Intrinsic value of fairness, diversity, novelty…
46. IRG
IRGroup @UAM
Rational and irrational bias in recommendation
The Information Retrieval General Reading Group
RMIT, June 3, 2021
Thank you!