Exploiting diffusion-based structured learning for item interactions representations in multimodal recommender systems

Abstract

Multimodal Recommender Systems (MRS) enhance the performance of recommendations by utilizing different item information, such as text, images, and audio. Existing non-graph-based MRS techniques combine embeddings (i.e., id and multimodal embedding) but ignore indirect and higher-order interactions. Graph-based MRS approaches use graph sparsification (GS) to construct item graphs and graph convolutional networks (GCNs) for higher-order interactions. However, GS reduces the item graph size, while GCNs ignore specific information due to their predefined weights. Hence, to mitigate the mentioned issues, this study proposes a Diffusion-based Structured Learning technique for Multimodal Recommender Systems (DSL-MRS) that improves the latent item graph information flow while maintaining its structure. Additionally, we used a graph attention neural network (GANN) to represent complex higher-order item-item interactions and implemented an attention mechanism to prioritize relevant nodes by assigning weights to neighbour. Also, for optimization, a Weighted Approximate-Rank pairwise (WARP) loss function has been used to prioritize predictions for observed items over those for unspecified items. To demonstrate the advantage of DSL-MRS, we conducted extensive experiments on three publicly available categories of Amazon datasets. The experimental findings showed that the proposed approach led to an average improvement of 5.8 % in R@20, 8.7 % in precision@20,7.8 % in NDCG@20 and 8.8 % in F-score@20 compared to the baseline model. Ablation studies demonstrate the value and efficacy of DSL-MRS, as its components degrade performance when removed.