引入模态自适应融合机制的多模态知识图谱推荐方法

Modality-Adaptive Fusion with Knowledge Graph Attention Network for Multi-Modal Recommendation

摘要: 知识图谱(KG)已广泛用于提升推荐系统性能，特别是在缓解数据稀疏问题方面。然而，现有多模态知识图谱推荐方法多采用固定或简单的融合策略(如拼接或线性加权)，难以充分建模不同模态对推荐决策的动态贡献，限制了异构信息的有效利用。为此，本文提出一种新颖的多模态自适应融合知识图谱注意力网络(MAF-KGAT)。该模型引入模态自适应门控机制，通过门控网络动态学习图像特征与结构嵌入的融合权重，实现更细粒度的模态融合。我们还结合预训练的CLIP图像编码器提取判别性视觉特征，并在KGAT架构基础上融合多模态信息以增强图谱传播能力。在AmazonBooks和MovieLens数据集上，MAF-KGAT在Recall@20和NDCG@20上分别提升了4.2%和3.5% (AmazonBooks)以及3.6%和2.9% (MovieLens)，尤其在冷启动场景下表现出更强的推荐能力，验证了其在个性化推荐中的有效性与鲁棒性。该方法已在电商和娱乐平台中验证，可有效缓解数据稀疏问题。

Abstract: Knowledge Graph (KG) has been widely used to improve the performance of recommendation systems, especially in alleviating the problem of data sparsity. However, most of the existing multimodal knowledge graph recommendation methods adopt fixed or simple fusion strategies (such as concatenation or linear weighting), making it difficult to fully model the dynamic contributions of different modalities to recommendation decisions and limiting the effective utilization of heterogeneous information. To this end, this paper proposes a novel multimodal adaptive fusion knowledge graph attention network (MAF-KGAT). This model introduces a modal adaptive gating mechanism. Through the gated network, it dynamically learns the fusion weights of image features and structural embeddings to achieve more fine-grained modal fusion. We also combine the pre-trained CLIP image encoder to extract discriminative visual features, and fuse multimodal information based on the KGAT architecture to enhance the graph propagation ability. Experiments on two real datasets, Amazonbook and MovieLens, show that MAF-KGAT significantly outperforms existing multimodal and knowledge graph recommendation methods in terms of metrics such as Recall and NDCG, especially demonstrating stronger recommendation capabilities in cold start scenarios. Its effectiveness and robustness in personalized recommendation have been verified.