GUIDE: Training Deep Graph Neural Networks via Guided Dropout Over Edges

Area of Science:

• Machine Learning
• Graph Neural Networks
• Artificial Intelligence

Background:

• Graph neural networks (GNNs) are powerful for graph-based semi-supervised learning (GSSL).
• A major limitation of existing GNNs is the oversmoothing issue, hindering their expressive power.
• Oversmoothing arises from excessive aggregation of information across different classes during node representation updates.

Purpose of the Study:

• To propose GUIded Dropout over Edges (GUIDE), an effective method for training deep GNNs.
• To alleviate the oversmoothing problem by reducing inter-class information aggregation.
• To enhance the discriminative ability of GNNs by promoting intra-class similarity and inter-class separation.

Main Methods:

• GUIDE removes inter-class edges based on edge strength, defined by their role in shortest paths.
• Edges with higher strength are more likely to connect nodes of different classes and are thus dropped.
• This strategy prioritizes retaining intra-class edges, strengthening within-class connections.

Main Results:

• Theoretical analysis supports GUIDE's effectiveness in mitigating oversmoothing.
• Experiments on six public benchmarks demonstrate GUIDE's consistent performance improvement.
• GUIDE outperforms state-of-the-art methods across various GNN backbones, both shallow and deep.

Conclusions:

• GUIDE effectively addresses the oversmoothing issue in GNNs.
• The method enhances node representations, leading to improved performance in GSSL tasks.
• GUIDE offers a robust solution for training deeper and more expressive GNNs.