Give Me MoE: Dynamic Expert Routing for Inductive Graph-based Recommendation

Abstract

Graph-based recommender systems capture high-order user–item dependencies but often underperform in cold-start settings, especially for new items with limited interactions. Inductive graph models generalize to unseen users and items via node features and neighborhood aggregation. However, most rely on shared parameters that may be suboptimal for heterogeneous neighborhoods, which can limit cold-start performance. We propose MoE-IGR, a Mixture-of-Experts module for inductive graph-based recommendation that replaces the shared post-aggregation update with dynamic, layer-wise expert routing, enabling context-dependent message passing. We instantiate MoE-IGR on PinSage and evaluate it on three real-world datasets under temporal, cold-user, cold-item, and joint cold-start protocols. MoE-IGR yields substantial gains (up to 16% Recall@20), largest in cold-item recommendation (up to 34.4% NDCG@20). Routing analysis reveals expert specialization: cold items focus on a subset of experts, warm items prefer others, and semantically related item groups exhibit similar routing patterns, suggesting that experts capture both collaborative and semantic signals.