Clarify Confused Nodes via Separated Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph neural networks (GNNs) excel at graph tasks but struggle with heterophilous nodes, violating the homophily assumption.
• Existing GNNs often use generic models or inefficient separate training, limiting performance on real-world graphs.

Purpose of the Study:

• To develop a novel metric and framework for effectively handling heterophilous nodes in GNNs.
• To improve GNN performance by enabling tailored processing of nodes based on their neighborhood characteristics.

Main Methods:

• Proposed a new metric, Neighborhood Confusion (NC), to reliably separate nodes into distinct groups.
• Introduced the Neighborhood Confusion-guided Graph Convolutional Network (NCGCN) framework, grouping nodes by NC values for intra-group weight sharing and message passing.

Main Results:

• Observed significant differences in intra-group accuracy and embeddings based on NC values.
• NCGCN demonstrated superior performance compared to state-of-the-art methods on both homophilous and heterophilous benchmarks.

Conclusions:

• Neighborhood Confusion provides an effective method for node separation in GNNs.
• The NCGCN framework significantly enhances GNN performance by addressing challenges posed by heterophilous nodes.