A Block-Based Adaptive Decoupling Framework for Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph neural networks (GNNs) excel at processing unstructured data through feature propagation.
• Feature propagation in GNNs can lead to over-smoothing, where node representations become overly similar with increased depth.
• This over-smoothing limits the effectiveness of deep GNN architectures.

Purpose of the Study:

• To propose a novel framework, Block-Based Adaptive Decoupling (BBAD), to mitigate over-smoothing in deep GNNs.
• To enhance the performance and parameter efficiency of GNNs by decoupling feature propagation and transformation.
• To provide mechanisms for automatic adjustment of propagation depth and flexible aggregation hops.

Main Methods:

• Developed a Block-Based Adaptive Decoupling (BBAD) framework utilizing backbone networks.
• Each block in the framework incorporates shallow GNNs with decoupled feature propagation and transformation.
• Introduced layer regularizations and flexible receptive fields for adaptive propagation depth and node-specific aggregation.

Main Results:

• Demonstrated that traditional coupled GNNs are more susceptible to over-smoothing in deep architectures.
• Showcased the diversity of outputs generated by different blocks within the BBAD framework.
• Achieved improved performance in semi-supervised and fully supervised node classification tasks on benchmark datasets.

Conclusions:

• The BBAD framework effectively addresses the over-smoothing problem in deep GNNs.
• The proposed method enhances the performance of various backbone GNN networks.
• BBAD outperforms existing deep GNNs in terms of performance and parameter efficiency.