Hypergraph Motif Representation Learning

Hypergraphs have emerged as a powerful tool for representing high-order connections in real-world complex systems. Similar to graphs, local structural patterns in hypergraphs, known as high-order motifs (h-motifs), play a crucial role in network dynamics and serve as fundamental building blocks across various domains. For this reason, predicting h-motifs can be highly beneficial in different fields. In this paper, we aim to advance our understanding of such complex high-order dynamics by introducing and formalizing the problem of h-motifs prediction. To address this task, we propose a novel solution that leverages both high-order and pairwise information by combining hypergraph and graph convolutions to capture hyperedges correlation within h-motifs, along with an innovative negative sampling approach designed to generate close-to-positive negative samples. To evaluate the effectiveness of our approach, we defined several baselines inspired by existing literature on hyperedge prediction methods. Our extensive experimental assessments demonstrate that our approach consistently outperforms all the considered baselines, showcasing its superior performance and robustness in predicting h-motifs.