作者
Yuankai Sun,Shengzhe Zhou,Yu Chen,Jinlan Zhao
摘要
Aiming at the problems of sparse user-item interaction data and insufficient utilization of modal information in traditional recommendation systems, this paper proposes a heterogeneous graph neural network recommendation algorithm that integrates multimodal features and graph structure learning. The TransMM-HGNR model realizes the deep integration of three types of modal information, namely visual, text and comment, through a unified framework to comprehensively improve the recommendation accuracy and robustness. Firstly, in the feature extraction stage, the pre-trained BERT and Sentence-BERT models are adopted to perform semantic encoding on user comments and item texts, and the VGG-16 network is utilized to extract the high-level visual features of the images, thereby obtaining a unified embedded representation of multi-source heterogeneous information. Secondly, the core module of the model, DyMAF, introduces a set of learnable query vectors combined with the Transformer attention mechanism. By adaptively calculating the interaction weights between modalities, it achieves dynamic fusion of features from different modalities. This mechanism can automatically capture the semantic correlations between modalities and suppress the interference of noise modalities, significantly enhancing the discriminability and generalization ability of feature representation. Ultimately, the fused item representation and user embedding achieve personalized recommendation through inner product calculation, and the training objective adopts the Bayesian personalized Ranking (BPR) loss function to optimize the ranking performance. The experimental part was evaluated on two public datasets, Amazon Baby and Sports, and compared with five baseline models such as LightGCN, VBPR, MMGCN, and DVBPR. The results show that TransMM-HGNR achieves the best performance in indicators such as Recall@20, respectively improving the accuracy of recommendations and the quality of rankings. Overall, this paper introduces the Transformer attention mechanism into the multimodal heterogeneous graph structure, achieving deep fusion of cross-modal semantic information and adaptive weight distribution, providing an efficient and scalable new approach for modeling recommendation systems in scenarios with sparse data and heterogeneous features.