GL-BKGNN: Graphlet-based Bi-Kernel Interpretable Graph Neural Networks

Abstract

While graph neural networks (GNNs) have successfully applied generalized convolution operations to the graph domain, providing a direct method to explain the dependency between the output and the presence of certain features and structural patterns in the input graph remains challenging. Inspired by image filters in standard convolutional neural networks (CNN), we propose a neural framework that connects bi-kernel with GNNs, incorporating predefined rules and focusing on the interpretability of graph filters during training. We address graph kernels based on their differentiability, using backpropagation of differentiable graph kernels for end-to-end training to generate gradient information for model optimization. Simultaneously, we leverage the stronger stability of non-differentiable graph kernels to capture local critical subgraphs, achieving deep fusion of structural features for node updates. Extensive experiments demonstrate that our model achieves competitive performance on graph classification datasets while providing additional benefits in interpretability.