Hyperbrokerage: A Machine Learning Approach to Brokerage in Hypergraphs

• In Event: Recent Advances in Social Network Analysis

Abstract

Brokerage has long been recognized as a cornerstone of social network analysis. Yet how to measure brokerage in different network settings remains an unsettled question and vibrant area of methodological innovation. With dramatic growth in the size, modality, and complexity of networks studied by social scientists over the past 20 years, methodologists have extended Gould and Fernandez’s (1989) time-tested typology for counting brokerage in simple, one-mode networks to two-mode (Jasny and Lubell 2015), dynamic (Spiro, Acton, and Butts 2013), and cultural networks (Leal 2025). However, when it comes to hypergraphs, researchers often rely on projection to simple networks for existing measures to work. In this paper, I introduce a machine learning approach to measure hyperbrokerage. Using simulated and empirical networks from a study of open science collaboration networks in the AI industry, I evaluate the validity of Gould and Fernandez’s (1989) operationalization in hypergraphs and show how projection-based measures (1) underestimate brokerage nested in community structure, (2) conflate structurally distinct roles such as gatekeepers and coordinators, and (3) scale poorly as network size increases. To address these limitations, I show that finding brokerage-related motifs in hypergraphs is a task equivalent to finding association rules among frequent itemsets in market basket analysis (Agrawal, Imieliński, and Swami 1993; Gondal 2022, 2025; Han, Kamber, and Pei 2011). Building on this equivalence, I develop a novel framework for measuring brokerage directly in hypergraphs that accounts for higher-order structure and avoid distortions by dyadic projection. I redefine Gould and Fernandez’s five types of brokerage for hypergraphs using association rule mining metrics such as support, lift, and conviction. I also show how these metrics can be combined with pruning of infrequent motifs to reduce computational costs in large-scale networks.

Author

• Zhuofan Li, Virginia Tech