Graph causal representation learning for out-of-distribution generalization.

Abstract

In recent years, graph neural networks (GNNs) have made remarkable achievements in graph representation learning, mainly by exploring the correlation between input graphs and labels. However, some studies have shown that GNN tends to predict the labels of graphs by utilizing spurious shortcut features rather than causally robust features. As a result, the existing model trained on the biased training dataset substantially declines its generalization ability. Therefore, this paper proposes a representation model based on causal analysis for out-of-distribution generalization. The main idea of the model is to utilize the graph attention mechanism as a parameterized node and edge mask generator to explicitly divide the input graph into the causal subgraph and the shortcut subgraph and then encode the corresponding disentanglement representation. On this basis, information theory is introduced to ensure the correct decoupling of representation to learn the perfect causal representation. Finally, the causal intervention is carried out at the representation level to reduce the correlation between the shortcut representation and the causal representation using the disentanglement representation. Experimental results on synthetic and real-world datasets demonstrate that our approach performs superior generalization over existing baselines.