Using Adversarial Autoencoders for Multi-Modal Automatic Playlist Continuation
1 like704 views
This document describes research using multi-modal adversarial autoencoders for automatic playlist continuation. Key points: - Adversarial autoencoders were used to encode playlist track information for recommendations. Additional metadata like titles was also included. - Experiments showed adversarial regularization improved performance over autoencoders alone. Including additional metadata also helped. - The best model used an adversarial autoencoder with 200 hidden units, trained for 20 epochs on playlists of 75,000 tracks including aggregated track metadata. This performed well on both development and challenge sets.
Using Adversarial Autoencoders for Multi-Modal Automatic Playlist Continuation
- 1. Iacopo Vagliano, ZBW Kiel Lukas Galke, University of Kiel Florian May, University of Kiel Ansgar Scherp, University of Stirling Using Adversarial Autoencoders for Multi-Modal Automatic Playlist Continuation ACM RecSys Challenge, 7 October 2018
- 2. www.moving-project.eu • Adversarial regularization improves autoencoders on images (Makhzani et al. 2015) • Adversarial autoencoders effective in recommendation tasks (Galke et al. 2018) • Smoothness on the code aids autoencoders to reconstruct highly sparse item vectors Motivation ACM RecSys Challenge, 7 October 2018 2 of 12 Makhzani, A. et al. (2015). “Adversarial Autoencoders”. In: CoRR abs/1511.05644. Galke, L. et al. (2018). Multi-Modal Adversarial Autoencoders for Recommendations of Citations and Subject Labels. ACM UMAP.
- 3. www.moving-project.eu • Adversarial regularization improves autoencoders on images (Makhzani et al. 2015) • Adversarial autoencoders effective in recommendation tasks (Galke et al. 2018) • Smoothness on the code aids autoencoders to reconstruct highly sparse item vectors Motivation ACM RecSys Challenge, 7 October 2018 • Are adversarial autoencoders also effective for automatic playlist continuation? • Is it beneficial aggregating item attributes (track title, album title, artist name)? Research Questions Makhzani, A. et al. (2015). “Adversarial Autoencoders”. In: CoRR abs/1511.05644. Galke, L. et al. (2018). Multi-Modal Adversarial Autoencoders for Recommendations of Citations and Subject Labels. ACM UMAP. 2 of 12
- 4. www.moving-project.eu • Set of m playlist P • Set of n tracks T • Sparse matrix X ϵ {0,1}m x n in the spanned space P x T • Xjk = 1 if the track k is in the playlist j (binary occurrence) Problem statement ACM RecSys Challenge, 7 October 2018 3 of 12
- 5. www.moving-project.eu • Multi-Modal Adversarial Autoencoders (AAE) • Train autoencoder on track sets (playlists) • Supply condition to the decoder (multi-modal) • Match code with a normal distribution for smooth representations (adversarial) Approach ACM RecSys Challenge, 7 October 2018 4 of 12
- 6. www.moving-project.eu • Multi-Modal Adversarial Autoencoders (AAE) • Train autoencoder on track sets (playlists) • Supply condition to the decoder (multi-modal) • Match code with a normal distribution for smooth representations (adversarial) Approach ACM RecSys Challenge, 7 October 2018 Bag of tracks 4 of 12
- 7. www.moving-project.eu • Multi-Modal Adversarial Autoencoders (AAE) • Train autoencoder on track sets (playlists) • Supply condition to the decoder (multi-modal) • Match code with a normal distribution for smooth representations (adversarial) Approach ACM RecSys Challenge, 7 October 2018 Playlist titles + aggregated track metadata Bag of tracks 4 of 12
- 8. www.moving-project.eu • Multi-Modal Adversarial Autoencoders (AAE) • Train autoencoder on track sets (playlists) • Supply condition to the decoder (multi-modal) • Match code with a normal distribution for smooth representations (adversarial) Approach ACM RecSys Challenge, 7 October 2018 Playlist titles + aggregated track metadata Bag of tracks Predicted tracks‘ probabilities 4 of 12
- 9. www.moving-project.eu • Multi-Modal Adversarial Autoencoders (AAE) • Train autoencoder on track sets (playlists) • Supply condition to the decoder (multi-modal) • Match code with a normal distribution for smooth representations (adversarial) Approach ACM RecSys Challenge, 7 October 2018 Autoencoder 4 of 12
- 10. www.moving-project.eu • Multi-Modal Adversarial Autoencoders (AAE) • Train autoencoder on track sets (playlists) • Supply condition to the decoder (multi-modal) • Match code with a normal distribution for smooth representations (adversarial) Approach ACM RecSys Challenge, 7 October 2018 MLP Autoencoder 4 of 12
- 11. www.moving-project.eu • Multi-Modal Adversarial Autoencoders (AAE) • Train autoencoder on track sets (playlists) • Supply condition to the decoder (multi-modal) • Match code with a normal distribution for smooth representations (adversarial) Approach ACM RecSys Challenge, 7 October 2018 MLP Autoencoder Adversarial Regularization 4 of 12
- 12. www.moving-project.eu 0 1 0 Example ACM RecSys Challenge, 7 October 2018 Workout Rock Chill Bag of tracks for “Walk of Life“ 5 of 12
- 13. www.moving-project.eu 0 1 0 Example ACM RecSys Challenge, 7 October 2018 Workout Rock Chill 1 0 1 1 1 … Workout Walk of life Bag of tracks for “Walk of Life“ Bag of words for “Workout“ 5 of 12
- 14. www.moving-project.eu 0 1 0 Example ACM RecSys Challenge, 7 October 2018 Workout Rock Chill 1 0 1 1 1 … Workout Walk of life Bag of tracks for “Walk of Life“ Bag of words for “Workout“ p(“We are the champions” |“Workout”) 5 of 12
- 15. www.moving-project.eu • Preliminary experiments • AAE optimization on a development set • Final experiments on the challenge set Experimental Procedure ACM RecSys Challenge, 7 October 2018 6 of 12
- 16. www.moving-project.eu • Goals • verify that the approach is effective • check if using additional metadata is beneficial • Comparison with 4 state-of-the-art methods • Run every methods with and without playlist titles • New user settings Preliminary experiments ACM RecSys Challenge, 7 October 2018 7 of 12 Test Training Usual split Our split
- 17. www.moving-project.eu • Adversarial regularization consistently improves the performance of autoencoders for automatic playlist continuation • Using playlist titles is beneficial Results of Preliminary Experiments ACM RecSys Challenge, 7 October 2018 Method MRR No Titles Titles IC 0.0515 (0.1700) - SVD 0.0658 (0.1946) 0.0662 (0.1953) AE 0.0645 (0.1855) 0.0679 (0.1913) AAE 0.0682 (0.1937) 0.0700 (0.1958) MLP - 0.0300 (0.1310) 8 of 12
- 18. www.moving-project.eu • Test 20 configurations 1. Different values of hidden units, epochs and code size with ntracks = 50,000 2. Choice of best-performing values while varying ntracks • Run every configuration with playlist titles only and with playlist titles + track metadata • Best performing configuration • ntracks = 75,000, 200 hidden units, 20 epochs, code size = 100 AAE optimization on the development set ACM RecSys Challenge, 7 October 2018 9 of 12 Hyperparameters Values ntracks 25 k, 50 k, 75 k, 100 k Hidden units 50, 100, 200 Epochs 10, 20 Code size 50, 100
- 19. www.moving-project.eu • Test of several configurations varying ntracks • Setting other parameters to best-performing on the dev set • 200 hidden units • 20 epochs • code size = 100 • Only considering aggregated metadata Final experiments on the challenge set ACM RecSys Challenge, 7 October 2018 10 of 12
- 20. www.moving-project.eu • Using aggregated metadata is beneficial Results on the Dev and Challenge Set ACM RecSys Challenge, 7 October 2018 Set R-Prec NDCG Clicks Titles Aggr. Titles Aggr. Titles Aggr. Dev 0.1063 0.1205 0.2092 0.2319 9.9477 7.9350 Challenge - 0.1787 - 0.3201 - 5.3510 11 of 12
- 21. www.moving-project.eu • Adversarial Autoencoders are effective for automatic playlist continuation • Aggregating items attributes is beneficial Conclusions ACM RecSys Challenge, 7 October 2018 12 of 12
- 22. www.moving-project.eu • Adversarial Autoencoders are effective for automatic playlist continuation • Aggregating items attributes is beneficial Conclusions ACM RecSys Challenge, 7 October 2018 MOVING is funded by the EU Horizon 2020 Programme under the project number INSO-4-2015: 693092 i.vagliano@zbw.eu @maponaso Code at https://github.com/lgalke/mpd-aae-recommender 12 of 12
- 23. www.moving-project.eu • 100 hidden layers • ReLU activation function • Drop probabilities after each layer 0.2 • Code size 50 • Adam for optimization • Initial learning rate 0.001 • Gaussian prior distribution Parameters on preliminary experiments ACM RecSys Challenge, 7 October 2018
- 24. www.moving-project.eu • 10,000 random playlists • 2,000 without title and either five tracks or ten tracks retained at random (0.5 probability) • 8,000 with a randomly selected number of retained tracks • either 100 or 25 tracks with a 0.2 probability each • either zero, one, five, or ten tracks, with probability 0.1 each • Random sampling approach • resulting distribution of tracks slightly different from the challenge set • Selection of tracks always random • no distinction between selecting the first tracks or random tracks • Naive approach for playlists with few tracks • Cannot remove more tracks than available • Negligible effects Development set characteristics ACM RecSys Challenge, 7 October 2018
- 25. www.moving-project.eu • Vocabulary with the 50,000 most frequent distinct words from the metadata • playlist title, track title, artist name, and album title • Different values of hidden units, epochs and code size on a predefined vocabulary based on Google • ntracks = 50,000. • Best-performing values (200, 20 and 100) chosen while varying ntracks Hyperparameters on Dev set ACM RecSys Challenge, 7 October 2018 Hyperparameters Values ntracks 25 k, 50 k, 75 k, 100 k Hidden units 50, 100, 200 Epochs 10, 20 Code size 50, 100