Another Perspective of Over-Smoothing: Alleviating Semantic Over-Smoothing in Deep GNNs.

Area of Science:

• Machine Learning
• Graph Neural Networks

Background:

• Graph neural networks (GNNs) excel at graph-structured data analysis but deeper models suffer performance degradation.
• The "over-smoothing" issue causes indistinguishable node embeddings and corrupts semantic structures, hindering GNN development.

Purpose of the Study:

• Investigate the over-smoothing issue in deep GNNs.
• Propose a novel strategy to preserve semantic structures in deep GNNs, addressing limitations of existing methods.

Main Methods:

• Introduce a cluster-keeping sparse aggregation strategy for deep GNNs.
• This plug-and-play strategy uses weighted residual connections to redistribute node aggregation extents, preserving semantic structure.

Main Results:

• The proposed strategy effectively preserves semantic structures, mitigating the negative effects of over-smoothing.
• Experiments demonstrate performance comparable to state-of-the-art methods on deep GNN tasks.

Conclusions:

• The cluster-keeping sparse aggregation strategy offers a viable solution to the over-smoothing problem in deep GNNs.
• This approach enhances GNN performance by maintaining semantic integrity in deeper architectures.