Adaptive Propagation Graph Convolutional Network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Convolutional Networks (GCNs) use node operations and message passing for graph data inference.
• Key challenges in GCNs include designing differentiable exchange protocols and understanding complexity-accuracy tradeoffs.
• Existing GCN models often use fixed communication steps, limiting adaptability.

Purpose of the Study:

• To introduce an adaptive propagation mechanism for GCNs.
• To enable nodes to independently control their communication steps.
• To improve GCN performance and efficiency while managing computational complexity.

Main Methods:

• Proposed Adaptive Propagation GCN (AP-GCN) model.
• Incorporated a halting unit per node, inspired by adaptive computation time, to decide communication continuation.
• Investigated a regularization term to balance communication and accuracy.
• Evaluated AP-GCN on multiple benchmark datasets.

Main Results:

• AP-GCN achieved superior or similar results compared to state-of-the-art GCN models.
• The adaptive approach required only a small overhead in additional parameters.
• The halting mechanism allowed for efficient, node-specific communication management.
• The regularization term successfully enforced a communication-accuracy tradeoff.

Conclusions:

• AP-GCN offers an effective method for enhancing GCN performance through adaptive communication.
• The model demonstrates a favorable balance between computational complexity and predictive accuracy.
• The open-source release facilitates further research and application of adaptive GCNs.