Enhancing graph neural networks for self-explainable modeling: A causal perspective with multi-granularity receptive fields

Abstract

Self-explainable Graph Neural Networks (GNNs) provide explanations alongside their predictions, making the model transparent and facilitating their wide adoption in high-stakes tasks. Current studies on constructing such GNNs are limited by the single receptive field, resulting in the modeling of spurious correlations in self-explainable GNNs. To address this issue, this paper introduces a GNN model with incorporated multi-granularity receptive fields, capturing causal correlations during the model construction and providing explanations alongside its predictions. Specifically, we employ closeness matrices with multiple structural orders to construct multi-granularity receptive fields for the model. Subsequently, we design a model architecture with sliced channels to integrate representations learned from multiple receptive fields heuristically. Objective functions from a causal perspective are further designed to guide the optimization of the proposed model. Experiments conducted on five real-world datasets and one synthetic dataset demonstrate the superior performance of the proposed model. In terms of classification accuracy, compared to SOTA baseline, the proposed GNN achieves the improvement of 0.17%, 1.99%, 0.70%, 0.83%, and 0.78% on the real-world datasets MUTAG, PTC, PROTEINS, IMDB-M, and IMDB-B, respectively. Compared to three self-explainable baselines, qualitative and quantitative studies are conducted on MUTAG, PTC, PROTEINS, IMDB-M, IMDB-B, and the synthetic dataset Spurious-Motif. Experimental results confirm that the proposed model can accurately identify the essential substructures, such as $N{O}_2$ in the MUTAG dataset. Additionally, the proposed model assigns significant weights to the motif part and distinguishes it from the base part in the Spurious-Motif dataset, enhancing the accuracy of graph classification and the explanations of the predictions. The classifications along with explanations obtained with this approach align with human cognition and experience.