DWSSA: Alleviating over-smoothness for deep Graph Neural Networks.

Area of Science:

• Machine Learning
• Graph Neural Networks

Background:

• Deep Graph Neural Networks (GNNs) excel at graph tasks but face over-smoothness due to shallow networks and uniform weight aggregation.
• Over-smoothness limits knowledge capture and hinders performance in deep GNNs by overemphasizing node features and static structures.

Purpose of the Study:

• To propose a novel Dynamic Weighting Strategy with Structure Augmentation (DWSSA) to alleviate over-smoothness in deep GNNs.
• To enhance the performance of GNNs by addressing limitations in knowledge capture and graph structure utilization.

Main Methods:

• Implemented Fuzzy C-Means (FCM) for node clustering and fuzzy assignment.
• Devised a novel metric function for dynamic weight adjustment in aggregation.
• Introduced a Structure Augmentation (SA) step using a parallelable KNN algorithm to add meaningful connections.

Main Results:

• The proposed DWSSA method effectively alleviates over-smoothness in deep GNNs.
• DWSSA enhances GNN performance across eleven homophilous and heterophilous graph benchmarks.
• The strategy reduces noisy aggregations and better utilizes topological information.

Conclusions:

• DWSSA offers a robust solution to the over-smoothness problem in deep GNNs.
• The method improves the effectiveness of GNNs in various graph-related tasks.
• DWSSA demonstrates significant potential for advancing deep learning on graphs.