Document Type : Original/Review Paper

Authors

• Seyedeh Hamideh Erfani ¹
• Mohammad Javad Fadaeieslam ²
• Reza Mortazavi ³
• Mohammad Rahmanimanesh ⁴

¹ Faculty of Electrical and Computer Engineering, Semnan Univercity.

² Electrical and Computer Engineering Faculty, Semnan University

³ Damghan University

⁴ Faculty of Electrical and Computer Engineering, Semnan University

Abstract

Graph Neural Networks (GNNs) frequently exhibit limited generalization performance when applied to out-of-distribution (OOD) data, primarily due to spurious correlations acquired during the training process. This paper introduces a hybrid methodology that explicitly reduces statistical correlations within the node representation space to improve OOD generalization capabilities. The proposed approach combines three complementary components: the Hilbert-Schmidt Independence Criterion (HSIC) for reducing inter-dimensional dependencies, variance regularization for stabilizing network weights, and a parallel Variational Graph Auto-Encoder (VGAE) for learning label-independent representations. The method is evaluated on node classification tasks using the GOOD benchmark datasets under covariate and concept shift scenarios. Comprehensive experiments demonstrate that eliminating statistical correlations in node representations enables the proposed GNN model to achieve higher OOD test accuracy and also achieve performance GAP reductions of up to 35% on average between Identical and Independent Distribution (IID) and OOD settings.

Keywords

• Graph neural networks
• Out-of-distribution generalization
• Statistical independence
• Node classification
• Distribution shift

Main Subjects

• H.3. Artificial Intelligence