METRIC: Multiple preferences learning with refined item attributes for multimodal recommendation

Abstract

In recent years, there has been a burgeoning interest in multimodal recommender systems, which integrate various data types to achieve more personalized recommendations. Despite this, the effective incorporation of user preferences for multimodal data and the exploration of inherent semantic relationships between modalities still need to be explored. Prior research typically utilizes multimodal data to construct item graphs, often overlooking the nuanced details within the data. As a result, these studies fail to thoroughly examine the semantic relationships between items and user behavioral patterns. Our proposed approach, METRIC, addresses this gap by delving deeper into multimodal information. METRIC consists of two primary modules: the multiple preference modelling (MPM) module and the item semantic enhancement (ISE) module. The ISE module performs relational mining across multiple attributes, leveraging the semantic structural relationships inherent in items. In contrast, the MPM module enables users to articulate their preferences across different modalities and facilitates adaptive fusion through an attention mechanism. This approach not only improves precision in capturing user preferences and interests but also minimizes interference from varying modalities. Our extensive experiments on three benchmark datasets substantiate METRIC's superiority and the efficacy of its core components.