Should Graph Convolution Trust Neighbors? A Simple Causal Inference Method

Fuli Feng, Weiran Huang, Xin Xin, Xiangnan He, Tat-Seng Chua

Abstract

Graph Convolutional Network (GCN) is an emerging technique for information retrieval (IR) applications. While GCN assumes the homophily property of a graph, real-world graphs are never perfect: the local structure of a node may contain discrepancy, e.g., the labels of a node's neighbors could vary. This pushes us to consider the discrepancy of local structure in GCN modeling. Existing work approaches this issue by introducing an additional module such as graph attention, which is expected to learn the contribution of each neighbor. However, such module may not work reliably as expected, especially when there lacks supervision signal, e.g., when the labeled data is small. Moreover, existing methods focus on modeling the nodes in the training data, and never consider the local structure discrepancy of testing nodes. This work focuses on the local structure discrepancy issue for testing nodes, which has received little scrutiny. From a novel perspective of causality, we investigate whether a GCN should trust the local structure of a testing node when predicting its label. To this end, we analyze the working mechanism of GCN with causal graph, estimating the causal effect of a node's local structure for the prediction. The idea is simple yet effective: given a trained GCN model, we first intervene the prediction by blocking the graph structure; we then compare the original prediction with the intervened prediction to assess the causal effect of the local structure on the prediction. Through this way, we can eliminate the impact of local structure discrepancy and make more accurate prediction. Extensive experiments on seven node classification datasets show that our method effectively enhances the inference stage of GCN.

Introduction: The Challenge of Local Structure Discrepancy

The fundamental assumption of Graph Convolutional Networks (GCNs) is homophily—the principle that connected nodes are likely to share similar properties. By aggregating features from a node's neighbors, GCNs generate powerful representations. However, real-world graphs are rarely perfectly homophilous. A node might be connected to "dissimilar" neighbors, a phenomenon known as local structure discrepancy. This discrepancy can lead the GCN to produce "poisoned" representations, degrading prediction accuracy.

Local structure discrepancy in the OGB-arXiv dataset, where node 2 has several cross-category edges.

While existing solutions like Graph Attention Networks (GATs) attempt to learn neighbor importance during training, they often struggle when labeled data is scarce. Furthermore, these methods do not explicitly address discrepancy at the inference stage for testing nodes, which may exhibit connection patterns different from the training set. This research asks a fundamental question: during inference, how can we determine if the local graph structure is beneficial or detrimental to the prediction?

Methodology: Causal GCN Inference (CGI)

To address this, the authors frame GCN prediction through the lens of causal inference. They model the prediction process as a causal graph where both the node's own features ($X$) and its neighbors ($N$) are causal factors for the final prediction ($\hat{Y}$). The proposed Causal GCN Inference (CGI) method introduces a test-time intervention strategy to mitigate the impact of noisy neighbors.

The causal framework for GCN prediction: (a) the general causal graph, (b) the original prediction process, and (c) the post-intervention prediction where neighbor influence is blocked.

CGI operates in three main steps:

  1. Causal Intervention: For a testing node, the method performs an intervention by "blocking" the neighbors (setting $N = \emptyset$). This forces the GCN to generate a "post-intervention" prediction ($\hat{y}^s$) based solely on the node's own features.
  2. Causal Uncertainty Estimation: To quantify the reliability of the local structure, the authors use a Monte Carlo approach. By randomly dropping edges and measuring prediction fluctuations, they calculate a "graph variance" score. High variance indicates an unreliable local structure.
  3. The Choice Model: A lightweight classifier (e.g., a Support Vector Machine) acts as a referee. It evaluates prediction confidence and causal uncertainty to decide whether to use the standard neighbor-aware prediction or the self-only post-intervention prediction.

Experimental Evaluation

The authors evaluated CGI across seven node classification datasets, including real-world citation networks (Cora, Citeseer, Pubmed, and OGB-arXiv) and synthetic benchmarks designed to simulate high discrepancy. For OGB-arXiv, they utilized high-quality RoBERTa embeddings to better isolate the impact of graph structure from feature quality.

Key findings include:

  • Robustness to Noise: On a synthetic Citeseer dataset where 50% of edges were misleading, CGI improved accuracy from 64.2% to 68.6%.
  • Consistent Gains: CGI consistently outperformed standard GCNs and simple ensemble methods across all real-world datasets, achieving a 5.5% relative improvement over neighbor-averaging ensembles on Cora.
  • Model Agnostic: The method was successfully applied to various state-of-the-art architectures, including APPNP, JKNet, and DAGNN, demonstrating its versatility as a "plug-and-play" inference enhancement.

Insights and Conclusion

CGI highlights the importance of adaptive locality at inference time. It treats the GCN as a black box and decides on-the-fly how to utilize neighbor information. A significant insight is the "Self-Trust" safety valve: when graph noise is high, relying on self-features is often superior to neighbor aggregation. Furthermore, the study suggests that as node feature quality improves (e.g., using RoBERTa), the relative importance of the graph structure may decrease, implying that GCNs sometimes compensate for weak feature engineering.

While CGI is currently limited to "conflict nodes"—where self-only and neighbor-aware predictions differ—it provides a robust, causal framework for improving GCN reliability in imperfect, real-world environments.

Summary updated

Abstract

Graph Convolutional Network (GCN) is an emerging technique for information retrieval (IR) applications. While GCN assumes the homophily property of a graph, real-world graphs are never perfect: the local structure of a node may contain discrepancy, e.g., the labels of a node's neighbors could vary. This pushes us to consider the discrepancy of local structure in GCN modeling. Existing work approaches this issue by introducing an additional module such as graph attention, which is expected to learn the contribution of each neighbor. However, such module may not work reliably as expected, especially when there lacks supervision signal, e.g., when the labeled data is small. Moreover, existing methods focus on modeling the nodes in the training data, and never consider the local structure discrepancy of testing nodes. This work focuses on the local structure discrepancy issue for testing nodes, which has received little scrutiny. From a novel perspective of causality, we investigate whether a GCN should trust the local structure of a testing node when predicting its label. To this end, we analyze the working mechanism of GCN with causal graph, estimating the causal effect of a node's local structure for the prediction. The idea is simple yet effective: given a trained GCN model, we first intervene the prediction by blocking the graph structure; we then compare the original prediction with the intervened prediction to assess the causal effect of the local structure on the prediction. Through this way, we can eliminate the impact of local structure discrepancy and make more accurate prediction. Extensive experiments on seven node classification datasets show that our method effectively enhances the inference stage of GCN.
BibTeX
@inproceedings{feng2021should,
  author = {Fuli Feng and Weiran Huang and Xin Xin and Xiangnan He and Tat-Seng Chua},
  title = {{Should Graph Convolution Trust Neighbors? A Simple Causal Inference Method}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2021},
  doi = {10.1145/3404835.3462971},
}
Figure