Reproduce “Disentangled Multimodal Representation Learning for Recommendation” on OpenResearch

About this paper

Fan Liu, Huilin Chen, Zhiyong Cheng, Anan Liu, Liqiang Nie, Mohan Kankanhalli

A novel recommendation model, Disentangled Multimodal Representation Learning (DMRL), profiles user preferences by disentangling factor-specific representations within each modality and modeling individual user attention to different modalities for each factor. This approach achieved superior recommendation performance across five Amazon review datasets compared to state-of-the-art baselines.

Key insights

• Explicitly disentangling factor-specific representations within each modality via distance correlation regularizer effectively avoids information redundancy and enhances the robustness of preference modeling.
• Modeling personalized user attention to different modalities for each item factor is crucial for optimal performance, as undifferentiated multimodal feature fusion can be detrimental.
• Leveraging comprehensive multimodal information, with proper disentanglement and attention, consistently yields superior recommendation quality compared to interaction-only or simpler multimodal fusion approaches.

Summary by alphaXiv.