Multi-modal sequential recommender system using mixture of experts

With the rise of recommender systems, sequential modeling has been widely adopted in practical personalized services. Most existing traditional methods rely mainly on item identifier sequences, which are insufficient to capture the semantic context behind user behaviors. In recent years, multi-modal sequential recommendation approaches have incorporated more signals such as textual and visual features to enhance user understanding. However, these methods still suffer from redundant information across modalities and a lack of explicit temporal modeling. To address these limitations, the Hierarchical Time-Aware Mixture of Experts for Multi-Modal Sequential Recommendation (HM4SR) framework offers an advanced and robust solution, which is systematically reproduced and ana lyzed in this dissertation. Specifically, HM4SR introduces an Interactive Mixture of Experts (IMoE) to remove redundant features and extract essential representations, and a Temporal Mixture of Experts (TMoE) to integrate time information for dynamic user interest learning. In addition, three auxiliary tasks—Category Prediction (CP), ID Contrastive Learning (IDCL), and Placeholder Contrastive Learning (PCL)—are designed to alleviate the data sparsity issue. Experiments on four public datasets show that HM4SR achieves notable performance gains over other advanced baselines, confirming its effectiveness and applicability.