Breaking the confinement of fixed nodes: A causality-guided adaptive and interpretable graph neural network architecture

Abstract

Graph neural networks (GNNs) have significantly advanced the processing of graph-structured data, where objects exhibit complex relationships and interdependencies. The graph convolutional network (GCN), as a representative technology, enables end-to-end learning of such data. However, as GNN technology continues to evolve, certain entrenched research paradigms have created bottlenecks in further development. A key issue arises from the common practice of predefining the graph structure, such as fixing the node degrees before learning the underlying graph structure. While this approach is often employed to constrain the learning process, it does not guarantee the optimal discovery of the graph’s potential structure. Specifically, the fixed node degree can limit the adaptability of the neighborhood, thereby influencing the model’s performance. In this paper, we provide an in-depth analysis of this limitation. From a theoretical perspective, we rigorously examine the constraints of traditional GNN architectures and highlight the importance of considering the dynamic relationship between input features and node degrees. Furthermore, we propose an optimization strategy for GNN learning architectures, utilizing causal inference techniques, and introduce an enhanced model, termed Causality-guided Graph Neural Network (C-GNN). Our theoretical contributions are supported by experimental validation, where comprehensive quantitative and qualitative evaluations demonstrate the superiority of the C-GNN model over traditional GNN architectures.