Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Prospective Preference Enhanced Mixed Attentive Model for Session-based Recommendation

Bo Peng, Chang-Yu Tai, Srinivasan Parthasarathy, Xia Ning

TL;DR

This work introduces P^2MAM, a session-based recommender that jointly models temporal patterns with a light-weight position-sensitive attention and estimates users' prospective preferences to guide attention. The model yields three scoring variants (O, P, and O-P) that respectively leverage temporal signals, prospective signals, or their combination, and demonstrates up to 19.2% improvement over state-of-the-art baselines across six datasets, along with substantial runtime speedups (47.7x on average) during testing. Ablation studies confirm the importance of position embeddings and predictive prospective preferences, while analysis of attention weights and cosine similarities shows the model learns meaningful, predictive representations. The approach also outperforms GRU- and graph-based methods, highlighting the effectiveness of direct sequence modeling with informed attention in sparse, real-world data.

Abstract

Session-based recommendation aims to generate recommendations for the next item of users' interest based on a given session. In this manuscript, we develop prospective preference enhanced mixed attentive model (P2MAM) to generate session-based recommendations using two important factors: temporal patterns and estimates of users' prospective preferences. Unlike existing methods, P2MAM models the temporal patterns using a light-weight while effective position-sensitive attention mechanism. In P2MAM, we also leverage the estimate of users' prospective preferences to signify important items, and generate better recommendations. Our experimental results demonstrate that P2MAM models significantly outperform the state-of-the-art methods in six benchmark datasets, with an improvement as much as 19.2%. In addition, our run-time performance comparison demonstrates that during testing, P2MAM models are much more efficient than the best baseline method, with a significant average speedup of 47.7 folds.

Prospective Preference Enhanced Mixed Attentive Model for Session-based Recommendation

TL;DR

This work introduces P^2MAM, a session-based recommender that jointly models temporal patterns with a light-weight position-sensitive attention and estimates users' prospective preferences to guide attention. The model yields three scoring variants (O, P, and O-P) that respectively leverage temporal signals, prospective signals, or their combination, and demonstrates up to 19.2% improvement over state-of-the-art baselines across six datasets, along with substantial runtime speedups (47.7x on average) during testing. Ablation studies confirm the importance of position embeddings and predictive prospective preferences, while analysis of attention weights and cosine similarities shows the model learns meaningful, predictive representations. The approach also outperforms GRU- and graph-based methods, highlighting the effectiveness of direct sequence modeling with informed attention in sparse, real-world data.

Abstract

Session-based recommendation aims to generate recommendations for the next item of users' interest based on a given session. In this manuscript, we develop prospective preference enhanced mixed attentive model (P2MAM) to generate session-based recommendations using two important factors: temporal patterns and estimates of users' prospective preferences. Unlike existing methods, P2MAM models the temporal patterns using a light-weight while effective position-sensitive attention mechanism. In P2MAM, we also leverage the estimate of users' prospective preferences to signify important items, and generate better recommendations. Our experimental results demonstrate that P2MAM models significantly outperform the state-of-the-art methods in six benchmark datasets, with an improvement as much as 19.2%. In addition, our run-time performance comparison demonstrates that during testing, P2MAM models are much more efficient than the best baseline method, with a significant average speedup of 47.7 folds.
Paper Structure (34 sections, 8 equations, 3 figures, 9 tables)

This paper contains 34 sections, 8 equations, 3 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Session-based Recommendation
  4. Sequential Recommendation
  5. Definitions and Notations
  6. Methods
  7. Session Representation ($\mathop{\mathtt{SR}}\limits$)
  8. Position-sensitive Preference Prediction ($\mathop{\mathtt{PS}}\limits$)
  9. Prospective Preference Enhanced Preference Prediction ($\mathop{\mathtt{P^2E}}\limits$)
  10. Recommendation Scores in $\mathop{\mathtt{P^2MAM}}\limits$
  11. Scores based on temporal patterns ($\mathop{\mathtt{P^2MAM\text{-}O}}\limits$)
  12. Scores based on users' prospective preferences ($\mathop{\mathtt{P^2MAM\text{-}P}}\limits$)
  13. Scores based on temporal patterns and users' prospective preferences ($\mathop{\mathtt{P^2MAM\text{-}O\text{-}P}}\limits$)
  14. Network Training
  15. Materials
  16. ...and 19 more sections

Figures (3)

  • Figure 1: Overall Architecture
  • Figure 2: Attention Weights from $\mathop{\mathtt{P^2MAM\text{-}O\text{-}P}}\limits$
  • Figure 3: Parameter Study