Exploiting heterogeneous information isolation and multi-view aggregation for multimodal recommendation

Recent research in multimodal recommendation has gravitated towards employing graph convolutional networks (GCNs) based propagation for generating user-item representations. However, there are two significant limitations in these research. Firstly, they usually fail to fully aggregate user-item modality and behavioral embedding information. This results in recommender systems getting recommendations that lack modality specificity or ignore raw interaction data. Secondly, existing models use summation to process heterogeneous information, which ignores differences in heterogeneous information. This manipulation leads to ambiguity in personalized recommendations and further exacerbates the interpretability challenges associated with GCNs. We propose a novel approach for solving the above problem by exploiting Heterogeneous information isolation and Multi-view aggregation for multimedia Recommendation (HMRec). HMRec uses isolated path aggregation to separately process the two types of heterogeneous information. Subsequently, it propagates this information through GCNs on three graphs: the user-item interaction graph, the user-user co-occurrence graph, and the item-item affinity graph. This multi-view propagation aims to capture optimal node representations. At the final stage of the model, the two kinds of information are calculated in isolation to obtain their respective recommendation lists. They are fused using the idea of integration to finally get a comprehensive recommendation result. Extensive experiments on three real-world datasets have demonstrated the superior performance of our proposed methods compared to state-of-the-art approaches.