Updating KernelCanvas for weightless graph classification

Raul Bezerra Barbosa, Diego Leonel Cadette Dutra, Priscila Machado Vieira Lima, Diego Carvalho, Felipe Maia Galvão França

Abstract: In a landscape where graph neural networks establish themselves as a standard in graph learning, we introduce a contrasting weightless graph learning architecture. We propose a novel representation for shallow graph learning based on weightless neural networks (WNN) and extend current WNN quantization techniques. Additionally, we assess this new architecture across widely adopted datasets. Our findings indicate that employing our representation and WNN models can yield superior results to those attained with extra-large graph neural networks or heavyweight kernel methods while imposing fewer demands on computer resources and memory. Furthermore, our results suggest that other classifiers can leverage the proposed representation for improved performance.