Modality-Guided Mixture of Graph Experts with Entropy-Triggered Routing for Multimodal Recommendation

Researchers introduce MAGNET, a new AI system for multimodal recommendation that combines user behavior, visual, and textual data through specialized graph neural network experts. The system uses entropy-triggered routing to automatically balance different data types and improve recommendations for sparse datasets and long-tail items.

• →MAGNET addresses the challenge of conflicting multimodal signals in recommendation systems through specialized expert networks.
• →The system employs a dual-view graph learning module that augments interaction graphs with content-induced edges for better coverage.
• →Three distinct expert roles (dominant, balanced, complementary) enable more interpretable fusion of behavioral, visual, and textual data.
• →A two-stage entropy-weighting mechanism prevents expert collapse and stabilizes training through progressive specialization.
• →Extensive experiments show consistent improvements over existing baseline recommendation systems.