AI-Augmented Multimodal Recommender Systems: Adaptive Fusion of Reviews, Images, and User Behaviour

Le résumé fourni par la source

The explosive growth of web content has introduced diverse modalities such as textual reviews, product images, and behavioural logs, which can be leveraged by a "Recommender System (RS)" to enhance personalization. However, existing multimodal RS often process modalities independently or apply rigid fusion strategies, limiting their effectiveness in terms of relevance, interpretability, and robustness, particularly under cold-start and sparse-data conditions. Similarly, earlier multimodal methods enrich embeddings with textual or visual features but lack adaptive fusion and efficient context propagation. To address these limitations, this paper introduces a hybrid Multimodal RS that integrates a "Cross-Modal Attention (CMA)" mechanism with "Log-Gabor Filter (LGP)", thereby enabling adaptive modality fusion and efficient context enrichment. The proposed CMA-LGP model first captures shortterm user intent through a lightweight sequence encoder, while semantic features are extracted from aggregated reviews and representative images using compact text and vision encoders. A CMA module adaptively weighs different modalities based on user intent, ensuring context-sensitive fusion. To further enrich sparse items, a shallow user–item–attribute LGP mechanism incorporates co-view and co-purchase signals. For deployment efficiency, the framework integrates practical strategies such as precomputing item embeddings, adopting a two-stage retrieval and re-ranking pipeline, and enabling mixed-precision inference for reduced latency. Experimental evaluations on standard public benchmarks demonstrate that the proposed CMA-LGP approach consistently outperforms collaborative filtering, sequential baselines, and existing multimodal RS models in terms of HR@10 and NDCG@10.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.