MoGL: A mixture of heterogeneous experts for collaborative graph learning.

Abstract

Graph Neural Networks (GNNs) have demonstrated remarkable success, yet they often exhibit limitations in capturing the complex local heterogeneity inherent in real-world graphs. While the Mixture of Experts (MoE) paradigm offers a promising direction, existing models are frequently constrained by standalone gating mechanisms, which introduce an unfavorable trade-off between predictive accuracy and computational efficiency. To address these challenges, we introduce the Mixture of Heterogeneous Experts for Collaborative Graph Learning (MoGL), a novel framework built upon a synergistic dual-expert architecture. MoGL strategically combines a standard GNN, tasked with capturing global structural information, with a lightweight and purpose-built local expert, the Graph Kolmogorov-Arnold Isomorphism Network (GKAIN). Inspired by the principles of Kolmogorov-Arnold Networks (KANs), GKAIN is uniquely designed to excel at modeling intricate non-linear relationships within local graph neighborhoods. To orchestrate the collaboration between these experts, we propose an innovative confidence-based gating mechanism that dynamically assigns task weights by leveraging the representations learned by GKAIN. This is further enhanced by a collaborative training paradigm, which employs a unified loss function to facilitate synergistic knowledge transfer between the experts. Extensive experiments on a diverse range of benchmarks demonstrate that MoGL achieves state-of-the-art performance on both homogeneous and heterogeneous graphs, thereby validating the superiority of our proposed architecture. The source code is available at: https://github.com/Season111/MoGL.