Abstract—Visual features and user-item interaction graphs enable recommendation systems to model complex relationships and improve accuracy. We propose Heterogeneous Graph Recommender Network with Bayesian Personalized Ranking for Fashion (HGRN-BPR-F), a Graph Neural Network that places multimodal edge features, explicit ratings and review sentiments, directly on user-item interactions. Unlike prior work aggregating features at nodes, our edge-centric design preserves interaction-specific context. The graph data is extracted from a real-world dataset of Amazon Fashion and exhaustive preprocessing is done to include relevant features for both users and products. Model training employs the Bayesian Personalized Ranking (BPR) loss function, which optimizes the ranking between positive and negative user–item pairs, with hard negatives used to enhance learning effectiveness. Our model achieves AUC of 0.885 ±0.010 evaluated over three seeds, demonstrating strong ranking performance with a 10.2% improvement over the rating-only baseline. Ablation experiments further demonstrate that combining ratings and sentiment as edge features substantially outperforms single-modality and node-feature-based configurations, thereby validating our edge-centric multimodal architecture.
 
Keywords—multimodal recommendation, Bayesian Personalized Ranking (BPR) loss, cosine similarity, Graph Neural Network (GNN), heterogeneous graphs


Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).