Joint subgraph independence for graph out-of-distribution generalization

Haoran Yu, Weifeng Liu, Yingjie Wang, Bao-di Liu, Dapeng Tao, Dan Zeng, Honglong Chen

Abstract: Most of the existing graph neural networks suffer from severe performance degradation under distribution shifts. The fundamental reason for degradation is that the model tends to use incorrect associations between feature and label for prediction. It significantly impacts the model’s performance under distribution shifts. Most of the existing graph out-of-distribution generalization research only considers the independence between subgraph and external factors to eliminate directly incorrect associations while ignoring spurious correlations between subgraphs. However, the spurious correlation between subgraphs still indirectly generates incorrect associations between spurious subgraph and labels. At the same time, it also results in indirect incorrect associations between causal subgraph and environment affecting the accurate selection of the causal subgraph. These problems further affect the model’s generalization ability. To solve the above problem, we propose a joint subgraph independence method to jointly eliminate the correlation between subgraphs and the correlation with external factors. Specifically, except for the independence of subgraphs from external factors, we learn a set of sample weights to make the spurious subgraph and causal subgraph independent of each other. Through eliminating the correlation between subgraphs, we avoid indirectly incorrect associations and promote more accurate identification of subgraph structures. At the same time, the selection of subgraph structure and independence with external factors provide a guiding prior for decorrelation between subgraphs, avoiding the loss of information and difficulty in optimization caused by dense decorrelation between all node features. We conduct extensive experiments on different datasets and the experimental results demonstrate the effectiveness of our method.