Research on GNNs with stable learning.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Causal Inference

Background:

• Traditional Graph Neural Networks (GNNs) struggle with Out-of-Distribution (OOD) data due to reliance on independent and identically distributed assumptions.
• Distribution discrepancies in real-world scenarios lead to unreliable GNN predictions in unknown domains.

Purpose of the Study:

• To enhance the stability and reliability of GNN cross-site classification for OOD data.
• To develop a machine learning method for extracting genuine causal features and eliminating spurious ones.

Main Methods:

• Introduced a feature sample weighting decorrelation technique in the random Fourier transform space.
• Developed a Stable-GNN model with a constrained sampling weight gradient update algorithm.
• Theoretically proved the algorithm ensures loss decrease and remedies weight updating shortcomings.

Main Results:

• The Stable-GNN model significantly reduces prediction bias on unseen test distributions.
• Experimental results show Stable-GNN outperforms current state-of-the-art GNN models.
• The proposed method guarantees predictive performance on training distribution data.

Conclusions:

• Stable-GNN enhances GNN stability and predictive accuracy in OOD scenarios.
• The framework offers a flexible approach to strengthen existing GNN models.
• This work addresses critical limitations of GNNs in real-world, non-stationary environments.