Hypergraph Representation Learning with Weighted- and Clustering-Biased Random Walks

TL;DR

This paper introduces a new framework for hypergraph representation learning that improves node classification by capturing higher-order structures.

Contribution

The novel WCRW-MLP framework combines weighted and clustering-biased random walks with an MLP for better hypergraph embedding.

Findings

WCRW-MLP outperforms state-of-the-art baselines on real-world hypergraph benchmarks.

Modeling co-occurrence strength and local clustering enhances structural embedding quality.

Abstract

Hypergraphs are powerful tools for modeling complex systems because they naturally encode higher-order interactions. However, most existing hypergraph representation-learning methods still struggle to capture such high-order structures, particularly in heterogeneous hypergraphs, which results in suboptimal performance on structure-sensitive tasks such as node classification. This paper presents WCRW-MLP, a new framework that integrates a Weighted- and Clustering-Biased Random Walk (WCRW) with a multi-layer perceptron. WCRW extends second-order random walks by introducing node-pair co-occurrence weights and triadic-closure clustering bias, enabling the walk to favor structurally significant and locally cohesive regions of the hypergraph. The resulting walk sequences are processed with Skip-gram to obtain high-quality structural embeddings, which are then concatenated with node attributes and fed into an MLP for classification. Experiments on several real-world hypergraph benchmarks show that WCRW-MLP consistently surpasses state-of-the-art baselines, validating both the efficacy of the proposed biasing strategy and the overall framework. These results demonstrate that explicitly modeling co-occurrence strength and local clustering is crucial for effective hypergraph embedding.

Figures