Bayesian Neighborhood Adaptation for Graph Neural Networks.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Science

Background:

• The neighborhood scope is crucial for graph neural network (GNN) performance.
• Current methods for determining scope are time-consuming and biased.
• Adaptive scope determination for GNNs is an underexplored area.

Purpose of the Study:

• To develop an adaptive method for determining the neighborhood scope in GNNs.
• To improve GNN performance on both homophilic and heterophilic graphs.
• To address the limitations of current two-stage GNN training approaches.

Main Methods:

• Modeling GNN message-passing as a stochastic process using a beta process for neighborhood scope.
• Employing a Bayesian framework for simultaneous inference of scope and GNN parameters.
• Theoretical analysis of scope inference's impact on GNN expressivity.

Main Results:

• Scope inference was shown to improve GNN expressivity.
• The proposed method is compatible with various GNN variants.
• Achieved competitive or superior performance on node classification tasks for benchmark datasets.

Conclusions:

• The Bayesian framework offers an effective way to adaptively determine GNN neighborhood scopes.
• The method provides well-calibrated predictions and enhances GNN performance.
• This approach advances GNN capabilities for diverse graph structures.