融合检索增强与动静态兴趣建模的多模态序列推荐方法研究

Research on Multimodal Sequential Recommendation Method Fusing Retrieval Augmentation and Dynamic-Static Interest Modeling

摘要: 多媒体内容的快速增长给推荐系统带来新挑战。传统序列推荐过度依赖文本信息，难以充分利用视觉语义；仅依赖时序建模，易忽略用户稳定偏好与物品结构关系；现有可解释推荐缺乏事实支撑，可信度较低。为此，本文在RACL-KAL基础上提出多模态推荐模型MM-RACL-KAL。模型融合文本与图像信息增强物品语义表示，通过检索增强扩展用户行为序列；采用Transformer与GNN实现动静态偏好融合建模，并结合多模态对比学习提升跨模态表示一致性；引入知识锚定大模型，生成有事实依据的可解释推荐。在Amazon-Fashion和MovieLens-Poster数据集上的实验表明，该模型在推荐性能与解释质量上均优于现有方法，验证了其有效性与可扩展性。

Abstract: The rapid growth of multimedia content poses new challenges to recommender systems. Traditional sequential recommendation relies excessively on textual information, making it difficult to fully utilize visual semantics. It only depends on temporal modeling, which tends to ignore users’ stable preferences and item structural relationships. Existing explainable recommendation methods lack factual support and thus have low credibility. To address these issues, this paper proposes a multimodal recommendation model MM-RACL-KAL based on RACL-KAL. The model fuses textual and visual information to enhance item semantic representation and extends user behavior sequences via retrieval augmentation. It adopts Transformer and GNN to achieve dynamic-static preference fusion modeling, combined with multimodal contrastive learning to improve cross-modal representation consistency. A knowledge-anchored large model is introduced to generate explainable recommendations with factual basis. Experiments on Amazon-Fashion and MovieLens-Poster datasets demonstrate that the proposed model outperforms state-of-the-art methods in both recommendation performance and explanation quality, verifying its effectiveness and scalability.