Disentangled Multimodal Representation Learning for Recommendation

Many multimodal recommender systems have been proposed to exploit the rich side information associated with users or items (e.g., user reviews and item images) for learning better representations improve recommendation performance. Studies from psychology show that individual differences in utilization of various modalities organizing information. Therefore, a certain factor an (such as appearance xmlns:xlink="http://www.w3.org/1999/xlink">quality ), features different are varying importance user. However, existing methods ignore fact contribute differently towards user's preference on factors item. In light this, this paper, we propose novel xmlns:xlink="http://www.w3.org/1999/xlink">Disentangled Multimodal Representation Learning (DMRL) model, which can capture users' attention each modeling. particular, employ disentangled representation technique ensure modality independent other. A mechanism is then designed factor. Based estimated weights obtained by mechanism, make recommendations combining scores preferences target over modalities. Extensive evaluation five real-world datasets demonstrate superiority our method compared methods.