Elastic Graph Neural Networks for Session Recommendation

ACL ARR 2026 January Submission1041 Authors

Keywords: session recommendation, graph neural networks, local adaptivity, global adaptivity

Abstract:

Session-based recommendation aims to predict users' next actions from anonymous interaction sessions. Existing approaches commonly adopt L2-based smoothing to enforce global stability and convergence; however, such strategies often overlook fine-grained local variations within session sequences and fail to fully exploit the adaptive modeling capacity of graph neural networks (GNNs). To overcome these limitations, we propose SR-EGNN, an Elastic Graph Neural Network for session-based recommendation that explicitly models adaptivity at both local and global levels. Specifically, the proposed framework decomposes adaptivity into two complementary components: local adaptivity, which captures subtle and dynamic variations within individual sessions, and global adaptivity, which preserves overall smoothness and training stability across the session graph. By jointly modeling these two aspects, SR-EGNN achieves a more effective trade-off between local sensitivity and global consistency than conventional smoothing-based methods. Concretely, SR-EGNN first encodes session representations using GNNs equipped with multi-head attention, and then applies dedicated local and global smoothing mechanisms to refine the representations. Extensive experiments on two real-world benchmark datasets demonstrate that SR-EGNN consistently outperforms strong baselines, validating the effectiveness of the proposed elastic modeling framework. The implementation is publicly available at
https://gitcode.com/Zane507/SR-EGNN

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: Dialogue and Interactive Systems,Information Extraction,Information Retrieval and Text Mining,Computational Social Science and Cultural Analytics

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: serial number

Submission Number: 1041