Scholarlink

In the era of digital economy, recommendatory systems face the dual challenges of insufficient use of semantic information and limited generalization ability of long tail data. It is difficult for traditional collaborative filtering methods to excavate deep semantic associations of items, while existing sequential recommendation models have significant bottlenecks in heterogeneous feature fusion and multi-scale interest modeling. In this paper, a dual-way collaborative semantic recommendation framework DualCBR is proposed. First, the collaborative embedding based on ID-PCA and the semantic embedding driven by LLM are generated respectively through the heterogeneous feature coding layer to preserve the group behavior pattern and fine-grained attribute features. Secondly, a lightweight cross-modal interaction layer is designed to dynamically align synergistic signals and semantic signals using the bidirectional cross-attention mechanism to solve the feature spatial heterogeneity problem. Furthermore, the multi-scale hybrid encoder is proposed, and the advantages of Transformer’s global dependency capture and RNN’s local time series modeling are combined to realize the collaborative expression of users’ long-term and short-term interests. Experiments on Amazon Beauty and Fashion datasets show that DualCBR achieves improvements of 5.7%–11.3% in NDCG@10 and HIT@10 compared with SASRec and other baseline models, while its bidirectional cross-attention mechanism increases the recommendation hit rate for long-tail items by 176%. The effectiveness of the model in alleviating the data sparsity problem through collaborative-semantic dual-path fusion is verified. This study provides a new technical path for semantic-enhanced personalized recommendation.

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

DualCBR: Cross-Modal Collaborative Filtering with Bidirectional Alignment for Long-Tail Recommendation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xin Li,
  • Lei Zhao,
  • Dekai Zhang,
  • Dawei Zhao,
  • Lijuan Xu,
  • Chunhui Wang,
  • Fuqiang Yu

摘要

In the era of digital economy, recommendatory systems face the dual challenges of insufficient use of semantic information and limited generalization ability of long tail data. It is difficult for traditional collaborative filtering methods to excavate deep semantic associations of items, while existing sequential recommendation models have significant bottlenecks in heterogeneous feature fusion and multi-scale interest modeling. In this paper, a dual-way collaborative semantic recommendation framework DualCBR is proposed. First, the collaborative embedding based on ID-PCA and the semantic embedding driven by LLM are generated respectively through the heterogeneous feature coding layer to preserve the group behavior pattern and fine-grained attribute features. Secondly, a lightweight cross-modal interaction layer is designed to dynamically align synergistic signals and semantic signals using the bidirectional cross-attention mechanism to solve the feature spatial heterogeneity problem. Furthermore, the multi-scale hybrid encoder is proposed, and the advantages of Transformer’s global dependency capture and RNN’s local time series modeling are combined to realize the collaborative expression of users’ long-term and short-term interests. Experiments on Amazon Beauty and Fashion datasets show that DualCBR achieves improvements of 5.7%–11.3% in NDCG@10 and HIT@10 compared with SASRec and other baseline models, while its bidirectional cross-attention mechanism increases the recommendation hit rate for long-tail items by 176%. The effectiveness of the model in alleviating the data sparsity problem through collaborative-semantic dual-path fusion is verified. This study provides a new technical path for semantic-enhanced personalized recommendation.