Jump to Content

Should Graph Convolution Trust Neighbors? A Simple Causal Inference Method

Fuli Feng
Qifan Wang
Tat-seng Chua
Weiran Huang
Xiangnan He
Xin Xin
SIGIR 2021 (2021) (to appear)

Abstract

Graph Convolutional Network (GCN) is an emerging technique in information retrieval (IR), which performs representation learning over graph structure. Graphs in IR applications possess locally strongly varying structure, pushing us to account for the discrepancy between subgraph structures. Existing work achieves the target by introducing an additional module such as graph attention, which learns to adaptively adjust the contribution of neighbors to the target node representation. However, such module may not work well in practice, especially when the labeled data are small, since the testing nodes may exhibit subgraph structures largely different from the labeled nodes. This work explores to resolve the impact of subgraph structure discrepancy for the testing nodes, which has received little scrutiny. From a new perspective, we investigate whether GCN should trust the subgraph structure of a testing node when performing inference. The main advantage of leaving the training stage unchanged is generality — it can be applied to most GCNs and improve their inference accuracy. Given a trained GCN model, the idea is to make a counter-factual prediction by blocking the graph structure, i.e., forcing the model to use each node’s own features to predict its label. By comparing the original prediction with the counterfactual prediction, we can assess the trustworthiness of neighbor nodes. Furthermore, we explore graph uncertainty that measures how the prediction would vary with the changes on graph structure, and introduce edge dropout into the inference stage to estimate graph uncertainty. We conduct empirical studies on seven node classification datasets to validate the effectiveness of our methods.