Reproduce “Semantic Item Graph Enhancement for Multimodal Recommendation” on OpenResearch

About this paper

Xiaoxiong Zhang, Xin Zhou, Zhiwei Zeng, Dusit Niyato, Zhiqi Shen

The SIGER framework enhances multimodal recommendation by integrating collaborative signals into item-item semantic graphs and mitigating noise in raw modality features, leading to more accurate item representations. It achieves superior performance compared to state-of-the-art baselines, with up to 7.57% improvement in general recommendation and substantial gains in cold-start scenarios on benchmark datasets.

Key insights

• Explicitly integrating collaborative signals into modality-specific item-item graphs captures a more comprehensive understanding of item relationships, moving beyond superficial content similarities.
• Personalized perturbation, adapted based on the embedding modulus and combined with contrastive learning, effectively mitigates structural noise from raw multimodal features, yielding more robust representations.
• Leveraging behavioral representations as anchors to guide the alignment of modality-specific semantic representations, followed by aligning fused semantics with behavior, refines the integration of heterogeneous information for enhanced consistency.

Summary by alphaXiv.