Collaborative local-global context modeling for session-based recommendation

Weiyue Li

doi:10.1016/j.ipm.2025.104196

Return

Collaborative local-global context modeling for session-based recommendation

2025-06-06

PRE

DOI：10.1016/j.ipm.2025.104196

Original

Original request for help

Save

Abstract

En 中文

Session-based recommendation systems (SBRSs) predict the next item in a session by analyzing user interactions. While current methods emphasize sequential item relationships, they often overlook temporal information that highlights subtle shifts in user preferences. This gap can limit their ability to adapt to dynamic user behavior, and recent advances have yet to effectively integrate both sequential and non-sequential item transitions, which may lead to biased modeling. To address these limitations, this paper introduces Coase, a novel SBRS model that unifies local and global context modeling to capture fine-grained dynamic user preferences. Coase transforms session sequences into session star graphs, employing a Bi-Gated Graph Self-Attention Network for local context modeling, and introduces SudokuFormer to model time-aware sequential transitions within a global session context through disentangled attention and stable feature fusion. A triple attention mechanism is then utilized to fully integrate local and global contextual features. Comprehensive experiments conducted on four publicly available datasets demonstrate that Coase improves Recall by 1.71%-1.83%, Mean Reciprocal Rank (MRR) by 2.73%-2.80%, and Normalized Discounted Cumulative Gain (NDCG) by 2.32%-2.43% across the top 5, 10, 15, and 20 items. Ablation studies validate the framework and components of Coase, while additional analyses examine the effect of session length, and visualization studies illustrate diverse attention patterns. This research contributes a novel approach to SBRS, offering promising advancements in recommendation accuracy and user experience.

Keywords:

Session-based recommendation

Collaborative effect

Graph neural network

Disentangled attention