AKBR: Learning Adaptive Kernel-based Representations for Graph Classification

TL;DR

The paper addresses graph classification by tackling limitations of traditional R-convolution kernels, notably their neglect of substructure importance and lack of end-to-end learning. It introduces Adaptive Kernel-based Representations (AKBR), which uses a feature-channel attention mechanism to weight substructure invariants derived from Weisfeiler-Lehman and Shortest Path kernels, forming an adaptive kernel matrix whose rows serve as graph embeddings fed into an MLP classifier. This end-to-end framework enables joint optimization of substructure weighting and classification, leading to superior performance over state-of-the-art graph kernels and several graph neural networks on standard benchmarks. The work demonstrates that capturing cross-graph patterns and substructure importance through learnable kernels can significantly boost graph classification accuracy and provides a versatile approach that can extend to other base kernels.

Abstract

In this paper, we propose a new model to learn Adaptive Kernel-based Representations (AKBR) for graph classification. Unlike state-of-the-art R-convolution graph kernels that are defined by merely counting any pair of isomorphic substructures between graphs and cannot provide an end-to-end learning mechanism for the classifier, the proposed AKBR approach aims to define an end-to-end representation learning model to construct an adaptive kernel matrix for graphs. To this end, we commence by leveraging a novel feature-channel attention mechanism to capture the interdependencies between different substructure invariants of original graphs. The proposed AKBR model can thus effectively identify the structural importance of different substructures, and compute the R-convolution kernel between pairwise graphs associated with the more significant substructures specified by their structural attentions. Since each row of the resulting kernel matrix can be theoretically seen as the embedding vector of a sample graph, the proposed AKBR model is able to directly employ the resulting kernel matrix as the graph feature matrix and input it into the classifier for classification (i.e., the SoftMax layer), naturally providing an end-to-end learning architecture between the kernel computation as well as the classifier. Experimental results show that the proposed AKBR model outperforms existing state-of-the-art graph kernels and deep learning methods on standard graph benchmarks.

Figures (4)

• Figure 1: The Framework of the Proposed AKBR Model.
• Figure 2: One Illustrative Instance of the Attention Mechanism for Feature Selection.
• Figure 3: The Attention Distribution for MUTAG.
• Figure 4: AKBR$_i$(WL) with Different Iterations