期刊: [Institute of Electrical and Electronics Engineers] 日期:2024-09-13卷期号:: 329-333
标识
DOI:10.1109/aiotc63215.2024.10748259
摘要
In recommender systems, analyzing user-item interactions history and capturing users' preferences is the starting point of all research. Multimodal information is also an important factor affecting users' decisions. Most previous studies have focused on using multimodal information as auxiliary information to model user-item interactions. [2] However, this method overlooks the rich content connections between users' or items' multimodal information. Therefore, we propose a recommendation method based on multimodal joint multi-graphs, referred to as FinalModel. Specifically, we utilize the interaction relationships between items and users to construct multimodal information of users, forming modality-aware latent semantic graph structures for both users and items. We employ graph convolution to explicitly inject higher-order affinity into the representations of users and items, obtaining multimodality-enhanced embedding expressions. Finally, these rich representations are integrated into the traditional collaborative filtering (CF) model. [3] Extensive experiments conducted on three real-world datasets demonstrate that our method outperforms the baseline models and advanced recommendation methods, validating the effectiveness of mining potential item-item and user-user relationships from multimodal features.