Beyond Message-Passing: Generalization of Graph Neural Networks via Feature Perturbation for Semi-Supervised Node...

Area of Science:

• Machine Learning
• Graph Neural Networks

Background:

• Graph neural networks (GNNs) are widely used in semi-supervised learning, with research focusing on effective graph filters and aggregation methods.
• Challenges arise from sparse training nodes and features (e.g., bag-of-words), leading to overfitting in projection matrices.

Purpose of the Study:

• To address the overfitting issue in GNNs caused by sparse node features.
• To propose an innovative perturbation technique to enhance GNN performance.

Main Methods:

• Introducing a novel perturbation technique that modifies initial features and the hyperplane.
• Increasing training variability to update all dimensions and reduce prediction variance.

Main Results:

• The proposed method significantly enhances node classification performance on real-world datasets.
• Achieved improvements of up to 46.5% in GNN algorithms.

Conclusions:

• This approach is the first to tackle GNN overfitting stemming from sparse node features.
• The perturbation technique effectively mitigates overfitting and boosts classification accuracy.