MMSRec: Multi-Modal Session-Based Personalised Product Recommendation for E-Commerce

Abstract

Personalised recommendation systems have become an essential component of modern e-commerce platforms, significantly influencing user engagement, customer satisfaction, and sales performance. Conventional recommendation approaches primarily rely on long-term historical user behaviour or globally popular products, limiting their ability to capture dynamic user intent, session-level preferences, and regional purchasing patterns. This paper proposes MMSRec, a hybrid recommendation framework that integrates session-based user modelling, multi-modal product representation, and regional trend analysis to generate context-aware and personalised recommendations. A Transformer-based sequential learning module models short-term user intent from interaction sessions, while textual and visual product features are fused to enhance semantic item representations. Furthermore, regional popularity signals are incorporated to improve the geographic relevance of recommendations. The proposed framework is evaluated on the Amazon Product Reviews and YooChoose benchmark datasets using Recall@10, NDCG@10, and Mean Reciprocal Rank (MRR) as evaluation metrics. Experimental results demonstrate that MMSRec consistently outperforms several state-of-the-art recommendation models, achieving up to 0.77 Recall@10, 0.50 NDCG@10, and 0.45 MRR on the Amazon dataset, while also delivering superior performance on YooChoose. These findings demonstrate the effectiveness of combining sequential user behaviour, multi-modal product information, and regional contextual knowledge to improve recommendation accuracy and ranking quality in large-scale e-commerce environments.