Graph Neural Network-Based Link Prediction for Highly Imbalanced Network Data

Abstract The rapid advancement of graph neural networks (GNN) has revolutionized the integration of complex network and neural network models, fostering diverse engineering applications such as the classification of assembly parts and the prediction of customer preferences. GNN-based link prediction (LP), as a representative study, has shown utility in resource allocation and system performance evaluation, as evidenced in the existing literature. However, the inherent challenge of imbalanced network data (i.e., sparse network) from engineering domains poses obstacles to accurate link predictions due to the lack of training data of positive links, leading to biases and reduced performance. In contrast to existing studies that focus on enhancing model architecture, this paper concentrates on exploring preprocessing methods, e.g., the data undersampling strategies, and two post-processing methods for accuracy calculation based on 1) probability threshold-based labeling and 2) probability rank-based labeling, and studies their impacts on the prediction results with a given GNN-based LP model architecture. Real-world network data sourced from shared mobility systems and product market systems, which exhibit various levels of imbalance, are used as test cases. This study contributes to a better understanding of effective data undersampling strategies and post-processing methods for handling imbalanced network data when applying GNN-based LP to real-world complex systems.