With the development of multimedia applications, multimodal recommendations play an essential role, as they can leverage rich contexts beyond user and item interactions. Existing methods mainly use them to help learn ID features; however, there exist semantic gaps among multimodal content features and ID features. Directly using multimodal information as an auxiliary would lead to misalignment in items' and users' representations. In this paper, we first systematically investigate the misalignment issue in multimodal recommendations, and propose a solution named AlignRec. In AlignRec, the recommendation objective is decomposed into three alignments, namely alignment within contents, alignment between content and categorical ID, and alignment between users and items. Each alignment is characterized by a distinct objective function. To effectively train AlignRec, we propose starting from pre-training the first alignment to obtain unified multimodal features and subsequently training the following two alignments together. As it is essential to analyze whether each multimodal feature helps in training, we design three new classes of metrics to evaluate intermediate performance. Our extensive experiments on three real-world datasets consistently verify the superiority of AlignRec compared to nine baselines. We also find that the multimodal features generated by our framework are better than currently used ones, which are to be open-sourced.
翻译:暂无翻译