Sequential Variational Autoencoders for Collaborative Filtering
0 likes156 views
The document discusses sequential variational autoencoders (SVAE) for collaborative filtering, presenting a robust mathematical model and an encoder-decoder formulation for recommender systems. It highlights the capability of SVAE to utilize historical preferences to predict future ones, improving upon traditional variational autoencoders by accounting for sequential information. Evaluation results indicate improved accuracy metrics across multiple platforms compared to existing methods.
Sequential Variational Autoencoders for Collaborative Filtering
- 1. Sequential Variational Autoencoders for Collaborative Filtering Noveen Sachdeva*, Giuseppe Manco, Ettore Ritacco, Vikram Pudi *supported by SIGIR Student Travel Grant and Flipkart
- 2. Context Recommendation as Preference Reconstruction Input Encoder Code Output Decoder
- 3. Context Recommendation as Preference Reconstruction Input Encoder Code Output Decoder
- 4. Context Recommendation as Preference Reconstruction Input Encoder Code Output Decoder
- 5. Context Recommendation as Preference Reconstruction Input Encoder Code Output Decoder
- 6. Context Recommendation as Preference Reconstruction Input Encoder Code Output Decoder
- 7. Context Recommendation as Preference Reconstruction Input Encoder Code Output Decoder
- 8. Context Recommendation as Preference Reconstruction Input Encoder Code Output Decoder
- 9. Context Recommendation as Preference Reconstruction Input Encoder Code Output Decoder
- 10. Variational Autoencoders for Recommender Systems • Elegant formulation as an encoder-decoder problem • Robust Mathematical Model based on variational approximation of the data likelihood • Encoder/decoder as probabilistic functions implemented through NNs
- 11. • Elegant formulation as an encoder-decoder problem • Robust Mathematical Model based on variational approximation of the data likelihood • Encoder/decoder as probabilistic functions implemented through NNs • Disregards the sequential nature of preferences Variational Autoencoders for Recommender Systems
- 12. Sequential Variational Autoencoder • At each step, history is encoded and exploited to decode the next preference • Encoding shaped via a Recurrent Nework … Input Encoder Code Output Decoder
- 20. Evaluation • Global accuracy Model Movielens Netflix NDCG Recall Precision NDCG Recall Precision @10 @100 @10 @100 @10 @100 @10 @100 @10 @100 @10 @100 POP 4.24 11.30 3.41 22.15 3.98 3.14 3.95 7.61 1.29 12.02 4.26 3.95 BPR∗ 8.51 16.66 5.20 29.38 7.27 4.51 8.94 12.68 2.63 19.45 8.45 5.89 RVAE 7.76 14.02 4.11 23.82 6.39 3.96 4.19 9.46 1.87 17.00 3.87 4.30 FPMC 5.65 12.06 3.46 23.10 4.77 3.46 11.01 13.55 3.49 20.82 10.12 6.04 CASER 16.32 27.55 11.35 46.01 13.15 6.45 16.46 19.79 6.33 28.33 14.38 7.54 MVAE 11.69 23.01 9.12 41.43 9.02 5.39 16.15 22.19 7.47 32.72 13.94 8.31 SVAE 17.81 29.93 12.48 49.09 14.40 6.93 24.64 26.77 8.93 35.58 21.93 10.55 MVAE∗ 27.82 39.79 17.46 59.70 23.01 9.59 35.41 37.70 13.50 47.22 31.39 15.20
- 21. Evaluation • Accuracy v/History Length
- 22. Evaluation • Accuracy v/Prediction Lookahead