Kernel Matrix-based Spd Representation For Graph Learning

Ye Xiao, Junbin Gao

Abstract: Graph pooling is the most crucial step designed to handle node representation from graph data and then produce appropriate graph representation which can be used for graph-level learning tasks. Inspired by studies demonstrating the superiority of kernel matrix in capturing nonlinear information thus better interpreting data representation in computer vision tasks, this paper explores the feasibility of kernel matrix applied in graph classification tasks. To be specific, we investigate the Deep Kernel SPD as a graph pooling strategy, namely Kernel SPD graph pooling. The overall experimental results reveal that the Kernel SPD graph pooling outperforms the current state-of-the-art baseline graph pooling strategy on specific benchmark datasets and is at least comparable to the baseline graph pooling strategy on multiple benchmark datasets, enabling us to further consider the potential of the kernel matrix-based SPD graph representation.