Scholarlink

<p>Session-based Recommendation (SBR) aims to predict the next item a user will interact with by leveraging historical behaviors in anonymous sessions. Most existing methods, including GNN-based approaches, often overlook time-series information within sessions, which is critical for capturing evolving user intents. This limitation hinders the ability of models to accurately infer time shifts in user preferences. Furthermore, existing SBR methods primarily rely on the current session to model user preferences, neglecting to utilize cross-session information from other similar sessions that may exhibit aligned behavioral patterns. To address these issues, we propose DICSI, a novel method that jointly models dynamic intents and cross-session information. Our framework combines two views: (i) Graph View: The graph attention convolution learns and fuses item features at global sessions and current session graphs. (ii) Sequence View: Multi-level temporal convolutions extract user dynamic intent, fused via adaptive attention mechanisms. DICSI further identifies sessions with similar intents, leveraging their collaborative signals through extraction and integration to enhance predictive performance. Across three real-world benchmarks, DICSI achieves average improvements of 11.3% on <i>P@20</i> and 3.6% on <i>MRR@20</i> over the best-performing baselines, demonstrating superior effectiveness.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dynamic intent-aware and cross-session integration for session-based recommendation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月29日

查看全文
  • Keqin Jiang,
  • Anyu Pan,
  • Yuchen Jiang,
  • Shulin Cheng

摘要

Session-based Recommendation (SBR) aims to predict the next item a user will interact with by leveraging historical behaviors in anonymous sessions. Most existing methods, including GNN-based approaches, often overlook time-series information within sessions, which is critical for capturing evolving user intents. This limitation hinders the ability of models to accurately infer time shifts in user preferences. Furthermore, existing SBR methods primarily rely on the current session to model user preferences, neglecting to utilize cross-session information from other similar sessions that may exhibit aligned behavioral patterns. To address these issues, we propose DICSI, a novel method that jointly models dynamic intents and cross-session information. Our framework combines two views: (i) Graph View: The graph attention convolution learns and fuses item features at global sessions and current session graphs. (ii) Sequence View: Multi-level temporal convolutions extract user dynamic intent, fused via adaptive attention mechanisms. DICSI further identifies sessions with similar intents, leveraging their collaborative signals through extraction and integration to enhance predictive performance. Across three real-world benchmarks, DICSI achieves average improvements of 11.3% on P@20 and 3.6% on MRR@20 over the best-performing baselines, demonstrating superior effectiveness.