Sampling nodes and hyperedges via random walks on large hypergraphs

TL;DR

This paper explores efficient sampling methods for large hypergraphs using random walks, compares existing techniques, introduces a non-backtracking variant, and validates its effectiveness on empirical data.

Contribution

It extends random walk sampling methods to hypergraphs, introduces a non-backtracking variant, and demonstrates its practical utility on large real-world hypergraphs.

Findings

Higher-order random walk outperforms other methods in hypergraph sampling

Non-backtracking extension provides unbiased estimators for hypergraph properties

Sampling estimates align with known empirical hypergraph characteristics

Abstract

Hypergraphs provide a fundamental framework for representing complex systems involving interactions among three or more entities. As empirical hypergraphs grow in size, characterizing their structural properties becomes increasingly challenging due to computational complexity and, in some cases, restricted access to complete data, requiring efficient sampling methods. Random walks offer a practical approach to hypergraph sampling, as they rely solely on local neighborhood information from nodes and hyperedges. In this study, we investigate methods for simultaneously sampling nodes and hyperedges via random walks on large hypergraphs. First, we compare three existing random walks in the context of hypergraph sampling and identify an advantage of the so-called higher-order random walk. Second, by extending an established technique for graphs to the case of hypergraphs, we present a non-backtracking variant of the higher-order random walk. We derive theoretical results on estimators based on the non-backtracking higher-order random walk and validate them through numerical simulations on large empirical hypergraphs. Third, we apply the non-backtracking higher-order random walk to a large hypergraph of co-authorships indexed in the OpenAlex database, where full access to the data is not readily available. Despite the relatively small sample size, our estimates largely align with previous findings on author productivity, team size, and the prevalence of open-access publications. Our findings contribute to the development of analysis methods for large hypergraphs, offering insights into sampling strategies and estimation techniques applicable to real-world complex systems.