Edge directionality in graph neural network

Abstract

This dissertation presents an empirical study aimed at improving the performance and stability of graph neural networks (GNNs) in handling complex graph data through design and methodological enhancements to better utilize graph information. The research focuses on the role of edge directionality in GNNs, emphasizing its critical importance in capturing dependencies and information flow in real-world graphs. Traditional GNNs often treat graphs as undirected or process only single-direction edges, which results in the loss of directional information and limits model performance. In heterogeneous graphs, the diversity of nodes and edges, explicit directionality, and complex information pathways further complicate model design. To address these challenges, the study proposes a direction-enhanced method based on Graph SAGE pretraining. This method introduces a direction assignment mechanism and a joint training framework for forward and backward edges. By recording intermediate feature representations at each layer, the improved message-passing mechanism enables the model to complement information during both forward and backward training, thereby capturing directional information more comprehensively. The joint training framework allows the model to simultaneously learn from both forward and backward graphs through shared parameters and intermediate features. Experiments conducted on public datasets (Wisconsin, Cornell, Texas, Actor, and Questions) demonstrate the effectiveness of the proposed method. The results show that it outperforms traditional approaches in terms of accuracy and stability. These findings highlight the importance of incorporating directional information in GNNs, bridging the gap between theory and practice.