Random Walk with Restart for Automatic Playlist Continuation and Query-specific Adaptations
1. Random Walk with Restart for Automatic
Playlist Continuation and Query-specific
Adaptations
Team Radboud:
Timo van Niedek
timo@mediadistillery.com
Arjen P. de Vries
arjen@acm.org
RecSys ‘18, Vancouver, Canada
2.
3. Random Walk with Restart for Automatic
Playlist Continuation and Query-specific
Adaptations
Team Radboud:
Timo van Niedek
timo@mediadistillery.com
Arjen P. de Vries
arjen@acm.org
RecSys ‘18, Vancouver, Canada
4. Current Challenges of Automatic Playlist Continuation
1)Scalability
2)Playlist purpose
3)Popularity bias
6. Recent work: Pixie1
• Recommender system at Pinterest
• Highly scalable:
3+ billion items, 200+ million users
1) Eksombatchai, C., Jindal, P., Liu, J. Z., Liu, Y., Sharma, R., Sugnet, C., ... & Leskovec, J. (2018, April). Pixie: A System for
Recommending 3+ Billion Items to 200+ Million Users in Real-Time. In Proceedings of the 2018 World Wide Web Conference on World
Wide Web (pp. 1775-1784). International World Wide Web Conferences Steering Committee.
7. The Playlist Graph
• Neighbor sampling in constant time
wrt. number of users & items
• Assumption: playlists contain similar
tracks (user judgment)
• Nearby tracks in graph are similar
Timo van Niedek. 2018. Random Walk with Restart for Automatic Playlist Continuation and
Query-Specific Adaptations. Master’s thesis. Radboud University, Nijmegen.
8. Random Walk with Restart
• Query track → Random Playlist → Random
Track → …
• Restart probability:
probability of restarting after every step
• Track-level recommendations:
Sorted list of visit counts
• Only local neighborhood is needed:
scalable!
Timo van Niedek. 2018. Random Walk with Restart for Automatic Playlist Continuation and
Query-Specific Adaptations. Master’s thesis. Radboud University, Nijmegen.
9. Multiple Random Walks with Restart
• Playlist-level recommendations:
Repeat Random Walk with Restart for
every track in playlist
Timo van Niedek. 2018. Random Walk with Restart for Automatic Playlist Continuation and
Query-Specific Adaptations. Master’s thesis. Radboud University, Nijmegen.
10. Evaluation
• Validation set: sample of the Million Playlist Dataset
• 10.000 playlists following structure of challenge set
12. 2. Playlist Purpose
• Playlist purpose = a set of target characteristics
• High quality recommendations match these characteristics
• Prefiltering: remove playlists that do not match target characteristics
13. Measuring Target Characteristics
• Target characteristics can be measured from the query playlist
• Playlist Title
• Query Track Features
14. Prefiltering approaches
• Title-based: remove playlists whose titles do not match query
playlist
• Feature-based: prune graph based on distance between playlist
models using features from Spotify API:
Acousticness, danceability, count, energy, explicitness,
instrumentalness, liveness, speechiness, tempo, valence, release
year
16. Degree Pruning
• Eksombatchai et al. (2018)1 have shown that removing low-
quality edges of popular pins is beneficial
• Highly popular songs are often added to playlists where
they do not belong
• Pruning edges of popular songs reduces random walk drift
1) Eksombatchai, C., Jindal, P., Liu, J. Z., Liu, Y., Sharma, R., Sugnet, C., ... & Leskovec, J. (2018, April). Pixie: A System for
Recommending 3+ Billion Items to 200+ Million Users in Real-Time. In Proceedings of the 2018 World Wide Web Conference on World
Wide Web (pp. 1775-1784). International World Wide Web Conferences Steering Committee.
17. Degree Pruning
• The degree of all track nodes t is updated to
by removing the lowest-quality edges according to
where t = collection of track features
20. 3. Popularity Bias
• Major issue in many academic approaches1
1) D. Jannach, I. Kamehkhosh, and G. Bonnin, “Biases in automated music playlist generation: A comparison of next-track recommending
techniques,” in Proceedings of the 2016 Conference on User Modeling Adaptation and Personalization. ACM, 2016, pp. 281–285.
21. 3. Popularity Bias
• Major issue in many academic approaches1
• Our methods reduce popularity bias
1) D. Jannach, I. Kamehkhosh, and G. Bonnin, “Biases in automated music playlist generation: A comparison of next-track recommending
techniques,” in Proceedings of the 2016 Conference on User Modeling Adaptation and Personalization. ACM, 2016, pp. 281–285.
22. Summary
• Scalable to large graphs
• Flexibility of integrating evidence
• Reduction of popularity bias