Hypers2v: a framework for structural representation of nodes in hypergraphs

Hypergraphs, unlike simple networks, can represent complex relationships among nodes, making them valuable for real-world applications such as social interactions. Learning embedded representations of nodes translates network structures into simplified spaces, enabling machine learning techniques designed for vector data to be applied to network data. However, existing methods often overlook the structural aspects of hypergraphs. This research introduces HyperS2V, a novel node embedding approach that emphasizes structural similarity within hypergraphs. The method involves defining hyper-degrees to capture nodes’ structural properties, creating functions to measure structural similarity, and generating embeddings using a multi-scale random walk framework. Experiments on both synthetic and real-world networks demonstrate HyperS2V’s superior performance in terms of interpretability and applicability to downstream tasks.