Adaptive receptive field graph neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) are powerful tools for representation learning but suffer from performance degradation in deeper architectures due to the over-smoothing problem.
• Over-smoothing causes node representations to become indistinguishable, limiting the effectiveness of deep GNNs.

Purpose of the Study:

• To address the over-smoothing issue in deep GNNs.
• To propose an novel GNN architecture that maintains performance with increased depth.
• To enhance node classification accuracy by adaptively aggregating neighbor information.

Main Methods:

• Introduced an Adaptive Receptive Field Graph Neural Network (ADRP-GNN) that utilizes a monolayer graph convolution layer.
• Developed a Multi-hop Graph Convolution Network (MuGC) to capture multi-hop neighbor information within a single layer.
• Incorporated a Meta Learner for adaptive receptive field generation and a Backbone Network for enhanced learning capacity.

Main Results:

• The proposed ADRP-GNN effectively mitigates the over-smoothing issue without requiring deeper networks.
• Experiments on eight datasets showed accuracy improvements ranging from 0.52% to 6.88% on node classification tasks compared to state-of-the-art methods.
• The adaptive receptive field mechanism allows integration with existing GNN frameworks for diverse applications.

Conclusions:

• ADRP-GNN offers a viable solution to the over-smoothing problem in GNNs.
• The adaptive aggregation of neighbor information enhances representation learning and classification performance.
• This architecture provides a flexible and effective approach for various GNN-based tasks.