Exploring centrality measures and encoding variants for graph classification in hyperdimensional computing

Abstract

Graph classification plays a central role in many scientific disciplines. Among existing approaches, classical machine learning methods and graph neural networks have demonstrated robust performance, though often at the cost of substantial computational resources. In recent years, Hyperdimensional Computing (HDC) has emerged as an efficient and noise-resilient alternative, offering a lightweight architecture well suited to resource-constrained environments. However, a key challenge in applying HDC to graph classification lies in generating highdimensional representations that effectively encode the structural patterns and latent information inherent in graphs. This paper builds upon the GraphHD framework by exploring alternative node centrality metrics. In addition, we introduce GraphHD-Level and GraphHD-Order, two novel variants that incorporate centrality information through distinct encoding strategies. Experiments on benchmark datasets from cheminformatics and bioinformatics (PROTEINS, DD, ENZYMES, NCI1, PTC-FM, and MUTAG) demonstrate that the proposed methods achieve classification performance comparable to standard approaches, while significantly reducing encoding time.