Cross-modal contrastive learning for aspect-based recommendation

Knowledge-enhanced recommender systems with aspects have improved recommendation performance by better profiling user preferences. Existing models can be divided into graph-based and text-based depending on the type of external knowledge: knowledge graph and review text. Since each knowledge provides different information from the scope and detail of aspects, it is necessary to integrate them for modeling sophisticated aspect-level preferences. However, it is difficult to directly fuse the aspects defined on two types of knowledge because they are expressed in different latent spaces. To tackle this problem, we explore self-supervised learning on multi-modal data. Specifically, we propose a novel model called COntrastive learning with croSs-MOdal aSpects (COSMOS). To take the data imbalance between knowledge graph and review texts into consideration, we devise a cross-modal contrastive learning scheme, which generates multiple views of a user or an item based on inter-modal correlation. With the correlation between aspects, COSMOS captures the inherent dependency between graph and text data. The fine-grained aspect-level preference, which contains salient features (from review text) as well as general ones (from knowledge graph), leads to providing high-quality recommendation results, even if a user only has one of the two data. Experimental results on two datasets show that COSMOS outperforms state-of-the-art recommender systems.