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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Graph Neural Networks (GNNs) excel but struggle with local graph heterogeneity.
• Existing Mixture of Experts (MoE) models face accuracy-efficiency trade-offs due to gating mechanisms.

Purpose of the Study:

• Introduce Mixture of Heterogeneous Experts for Collaborative Graph Learning (MoGL) to address GNN limitations.
• Develop a novel framework combining global and local experts for superior graph learning.

Main Methods:

• Propose a dual-expert architecture: a standard GNN for global structure and Graph Kolmogorov-Arnold Isomorphism Network (GKAIN) for local patterns.
• Implement a confidence-based gating mechanism for dynamic expert weight assignment.
• Utilize a collaborative training paradigm with a unified loss function for synergistic knowledge transfer.

Main Results:

• MoGL achieves state-of-the-art performance on diverse benchmarks for both homogeneous and heterogeneous graphs.
• Demonstrates superior predictive accuracy and computational efficiency compared to existing methods.
• Validates the effectiveness of the synergistic dual-expert architecture and confidence-based gating.

Conclusions:

• MoGL offers a significant advancement in graph learning by effectively capturing local heterogeneity.
• The proposed framework provides a scalable and efficient solution for complex graph data.
• MoGL represents a new direction for developing powerful and versatile graph representation learning models.