Multimodal Graph Learning and Sequential Modeling for Video Recommendation

The proliferation of multimedia content demands advanced video recommendation systems capable of capturing complex multimodal, heterogeneity, cold-start scenarios and dynamic user-item interactions. In this paper, we introduce a framework that integrates graph-based modeling, multimodal representation learning, and sequential modeling to address these challenges. Our approach begins with the transformation of a Heterogeneous Information Network (HIN) into a homogeneous video graph, preserving high-order semantic relationships through meta-path-guided transformations. We integrate multimodal features aligned into a shared embedding space via geometric encoding. To capture the temporal dynamics of user interactions, gated recurrent units are employed, providing a nuanced understanding of evolving preferences. The framework is optimized with a multi-objective loss to balance ranking accuracy and structural regularization. Experiments on two datasets demonstrate our framework superiority over state-of-the-art approaches. Additionally, our approach demonstrates exceptional resilience in addressing cold-start scenarios and provides explainable recommendations by revealing semantic connections between user history and suggested content.