These are the slides of my talk at the 2019 Netflix Workshop on Personalization, Recommendation and Search (PRS). This talk is based on previous talks on research we are doing at Spotify, but here I focus on the work we do on personalizing Spotify Home, with respect to success, intent & diversity. The link to the workshop is https://prs2019.splashthat.com/. This is research from various people at Spotify, and has been published at RecSys 2018, CIKM 2018 and WWW (The Web Conference) 2019.
3. Spotify’s mission is to
unlock the potential of
human creativity — by
giving a million creative
artists the opportunity
to live off their art and
billions of fans the
opportunity to enjoy
and be inspired by it.
7. Home
Home is the default screen of the mobile app
for all Spotify users worldwide.
It surfaces the best of what Spotify has to
offer, for every situation, personalized
playlists, new releases, old favorites, and
undiscovered gems.
Help users find something they are going to
enjoy listening to, quickly.
11. Streaming UserBaRT
Explore, Exploit, Explain: Personalizing Explainable Recommendations with Bandits. J McInerney, B Lacker, S Hansen, K Higley, H.Bouchard,
A Gruson & R Mehrotra. RecSys 2018.
BaRT: Machine learning algorithm for
Spotify Home
12. BaRT (Bandits for Recommendations as Treatments)
how to rank playlists (cards) in each shelf first, and then how to rank the shelves?
13. Explore vs Exploit
Flip a coin with given probability of tail
If head, pick best card in M according to predicted reward r → EXPLOIT
If tail, pick card from M at random → EXPLORE
BaRT: Multi-armed bandit algorithm for Home
https://hackernoon.com/reinforcement-learning-part-2-152fb510cc54
14. Success is captured by the reward function
Reward
Binarised Streaming Time
BaRT UserStreaming
success is when user
streams the playlist
for at least 30s
16. consumption time of a sleep playlist is longer than average playlist consumption time
jazz listeners consume Jazz and other playlists for longer period than average users
Is success the same for all playlists and users?
17. one reward function for
all users and all playlists
success independent of user and
playlist
one reward function per
user x playlist
success depends on user and
playlist
too granular, sparse, noisy, costly to
generate & maintain
one reward function per
group of users x playlists
success depends on group of users
listening to group of playlists
Personalizing the reward function
18. Co-clustering using streaming time
users
playlists
user
groups
playlist groups
Dhillon, Mallela & Modha, "Information-theoretic co-clustering”, KDD 2003.
group = cluster
group of user x playlist = co-cluster
21. co-clustering performs better than
baseline, user-only & playlist-only
clustering
global distribution-aware reward (no cluster)
better than baseline
mean better than additive itself better than
continuous
Experiments
… expected stream rate using distribution-aware reward functions
22. Towards personalizing with respect to success
Deriving User- and Content-specific Rewards for Contextual Bandits. P Dragone, R Mehrotra & M Lalmas. WWW 2019.
co-clustering and distribution-aware reward functions highlight importance of affinity type
features (type of content x type of user) over generic features (age/day)
24. User intents on Home
Passively Listening
- quickly access playlists or saved music (2)
- play music matching mood or activity (4)
- find music to play in background (6)
Actively Engaging
- discover new music to listen to now (3)
- to find X (5)
- save new music or follow new playlists for later (7)
- explore artists or albums more deeply (8)
Other
Home is default screen (1)
Mixed-methods approach: user interview & in-app survey & Bayesian non-parametric clustering
intent important to interpret user interaction
Jointly Leveraging Intent and Interaction Signals to Predict User Satisfaction with Slate Recommendations. R Mehrotra, M Lalmas, D Kenney,
T Lim-Meng & G Hashemian. WWW 2019.
25. Varying degree of satisfaction across intents
Passively Listening
(2) quickly access playlists or saved music
(4) play music matching mood or activity
(6) play music to play in background
Actively Engaging
(3) discover new music to listen to now
(5) to find X
(7) save new music or follow new playlists for later
(8) explore artists or albums more deeply
Other
(1) Home is default screen
some intents easier to satisfy
intent 6 → lean-back mood
intents 3 & 7 → discovery mood
26. Modeling user intents for Home
Global Model: single model
across all-intents
● ignores intent-level variations in
interaction data
● inadvertently suppresses
important variations
Per-intent Model: separate for
each intent
● using local intent specific
information only
● ignores information & insights
from other intent
Multi-level model: shared
learning across intents
● account for intent level grouping
● allows for variability in interaction
behavior across intents
● facilitate information sharing
across different intents
20% improvement in satisfaction
prediction over global model
shared learning across intents beneficial
27. different interaction important for different intents
(2) quickly access playlists or saved music → time to success and dwell time
(8) explore artists or albums more deeply → building relationships (save or download tracks)
Towards personalizing with respect to user intents
29. Relevance
playlist is relevant to user if closely resembles user interest profile
(embedding based representation of users & tracks)
Satisfaction
playlist satisfies user if user interacts with it
(stream for more than 30s)
Diversity
playlist is diverse if contains tracks from artists belonging to different groups
(popularity)
Towards a Fair Marketplace: Counterfactual Evaluation of the trade-off between Relevance, Fairness & Satisfaction in Recommendation Systems.
R Mehrotra, J McInerney, H Bouchard, M Lalmas & F Diaz. CIKM 2018.
One user intent on Home is “discover new music”
Spotify’s mission is to
unlock the potential of
human creativity — by
giving a million creative
artists the opportunity to
live off their art and
billions of fans the
opportunity to enjoy and
be inspired by it.
30. very few playlists have both high
relevance & high diversity
diversity relevance
The interplay between relevance, diversity & satisfaction
optimizing for relevance vs optimizing for
diversity and its impact on satisfaction
… but … but … but
Normalized Diversity Estimate
Match fans and artists in a personal and relevant way.
31. % loss in diversity % loss in relevance % gain in satisfaction
optimizing relevance (β=1) 69.1 0 0
Optimizing diversity (β=0) 0 57.7 -32.2
trade-off relevance & diversity (β =0.7) 42.7 9.8 -10.2
guaranteed relevance (rel >= 0.7) 51.7 7.8 4.4
diversity affinity aware 15 21.2 12.1
simple interpolation from β = 0 to 1 leads to high satisfaction loss or high diversity loss
guaranteed relevance hurts diversity
diversity affinity aware provides the best overall trade-off → personalizing diversity
Towards personalizing with respect to diversity