Scholarlink

<p>Sequential recommendation models built on self-attention have achieved strong performance, yet most existing approaches including SASRec and its variants remain limited by their reliance on static positional encodings and context-agnostic attention mechanisms. These models fail to account for rich contextual factors such as temporal irregularity, interaction type, session behavior, and recency, which strongly influence user intent. To address these limitations, we propose CSASRec, a novel context-enhanced self-attentive architecture that incorporates contextual information at multiple stages of the recommendation pipeline. CSASRec introduces three major components: Contextual Positional Encoding (CPE), which fuses positional cues with time-aware and behavior-aware signals; Context-Gated Self-Attention (CGSA), which integrates context gating directly into query–key–value projections to modulate attention relevance; and Multi-Level Context Refinement (MLCR), a hierarchical refinement mechanism that propagates contextual semantics across attention layers to strengthen both short-term and long-term user preference modeling. Experiments conducted on Amazon Beauty, Amazon Games, MovieLens-1M, and Steam datasets demonstrate that CSASRec consistently outperforms SASRec, BERT4Rec, GRU4Rec+, BSARec and LHASRec across Hit@10 and NDCG@10 metrics, with an average improvement of 7.13% and 9.49%, respectively. Ablation studies confirm the importance of each proposed component, with CGSA contributing the most significant uplift by enabling context-sensitive relevance estimation. The proposed model is efficient, interpretable, and scalable for modern large-scale sequential recommendation tasks.</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

A context enhanced self-attentive framework for sequential recommendation with multi level context modeling

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月21日

查看全文
  • Md Mahtab Alam,
  • Mumtaz Ahmed,
  • Pramit Kumar Samant,
  • Abdulatif Alabdulatif

摘要

Sequential recommendation models built on self-attention have achieved strong performance, yet most existing approaches including SASRec and its variants remain limited by their reliance on static positional encodings and context-agnostic attention mechanisms. These models fail to account for rich contextual factors such as temporal irregularity, interaction type, session behavior, and recency, which strongly influence user intent. To address these limitations, we propose CSASRec, a novel context-enhanced self-attentive architecture that incorporates contextual information at multiple stages of the recommendation pipeline. CSASRec introduces three major components: Contextual Positional Encoding (CPE), which fuses positional cues with time-aware and behavior-aware signals; Context-Gated Self-Attention (CGSA), which integrates context gating directly into query–key–value projections to modulate attention relevance; and Multi-Level Context Refinement (MLCR), a hierarchical refinement mechanism that propagates contextual semantics across attention layers to strengthen both short-term and long-term user preference modeling. Experiments conducted on Amazon Beauty, Amazon Games, MovieLens-1M, and Steam datasets demonstrate that CSASRec consistently outperforms SASRec, BERT4Rec, GRU4Rec+, BSARec and LHASRec across Hit@10 and NDCG@10 metrics, with an average improvement of 7.13% and 9.49%, respectively. Ablation studies confirm the importance of each proposed component, with CGSA contributing the most significant uplift by enabling context-sensitive relevance estimation. The proposed model is efficient, interpretable, and scalable for modern large-scale sequential recommendation tasks.