SES: Bridging the Gap Between Explainability and Prediction of Graph Neural Networks

TL;DR

SES addresses the gap between explainability and prediction in graph neural networks by introducing a self-explained and self-supervised framework. It couples a co-trained mask generator with a graph encoder during Explainable Training to produce $M_f$ and $M_s$, and then leverages these masks in Enhanced Predictive Learning to form mask-based positive/negative pairs and a triplet loss for representation learning. Across real-world and synthetic datasets, SES achieves state-of-the-art performance in both explanation quality (e.g., Fidelity+ and AUC metrics) and node classification accuracy, while offering faster explanation generation than post-hoc methods. The work demonstrates that integrating explainability directly into the training loop can enhance both interpretability and predictive power, enabling more reliable downstream use of GNNs.

Abstract

Despite the Graph Neural Networks' (GNNs) proficiency in analyzing graph data, achieving high-accuracy and interpretable predictions remains challenging. Existing GNN interpreters typically provide post-hoc explanations disjointed from GNNs' predictions, resulting in misrepresentations. Self-explainable GNNs offer built-in explanations during the training process. However, they cannot exploit the explanatory outcomes to augment prediction performance, and they fail to provide high-quality explanations of node features and require additional processes to generate explainable subgraphs, which is costly. To address the aforementioned limitations, we propose a self-explained and self-supervised graph neural network (SES) to bridge the gap between explainability and prediction. SES comprises two processes: explainable training and enhanced predictive learning. During explainable training, SES employs a global mask generator co-trained with a graph encoder and directly produces crucial structure and feature masks, reducing time consumption and providing node feature and subgraph explanations. In the enhanced predictive learning phase, mask-based positive-negative pairs are constructed utilizing the explanations to compute a triplet loss and enhance the node representations by contrastive learning.

Figures (8)

• Figure 1: Categories of GNNs from explanation and prediction perspectives.
• Figure 2: The framework of SES involves two phases: explainable training and enhanced predictive learning. The parameters of the graph encoder are shared in two phases. The green dashed arrows marked the process where the mask generator and the graph data are fed into the graph encoder.
• Figure 3: The framework of the mask generator in SES.
• Figure 4: Parameter sensitivities of SES.
• Figure 5: Visualization of node representations after training on Citeceer.