Reproduce “Enhancing Dyadic Relations with Homogeneous Graphs for Multimodal Recommendation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hongyu Zhou, Xin Zhou, Lingzi Zhang, Zhiqi Shen

The DRAGON model enhances multimodal recommendation by simultaneously learning user and item representations from both user-item interaction graphs and newly constructed homogeneous graphs capturing internal user-user and item-item relations, while employing an attentive concatenation strategy for effective multimodal feature fusion. It consistently outperformed strong baselines, achieving up to 32.11% improvement in Recall@10 on Amazon datasets and demonstrating competitive scalability on a large dataset.

Key insights

  • Simultaneously learning dual representations that incorporate both user-item inter-relations and user-user/item-item intra-relations significantly enhances recommendation accuracy.
  • Standard multimodal fusion techniques, such as pooling or attention-based summation, can degrade performance by not effectively combining complementary information.
  • Explicitly modeling high-order internal relationships using dedicated user co-occurrence and item semantic homogeneous graphs improves the richness of user and item embeddings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.