Beyond Pairwise Network Interactions: Implications for Information Centrality
提出用超图直接建模群体交互，通过两步扩散过程计算中心性，在开源软件、高中互动和金融共现三个场景中比传统成对网络更准确预测项目成功、学生受欢迎度和同日回报，并发布Python工具包。
Organizations and policymakers increasingly rely on network metrics to decide whom to inform, monitor, or support. Yet most networks treat interactions as pairs, even when the underlying activity occurs in groups—project teams, chat channels, meetings, or news articles that mention multiple firms. Collapsing groups into one-to-one links can misidentify who matters. We propose a practical alternative: Model group interactions directly as a hypergraph and compute centrality from a two-step diffusion process that captures how information moves across and within groups. The approach provides a transparent way to incorporate domain knowledge (e.g., whether people enter large or small groups first) and produces testable interpretable rankings. We evaluate the method in three settings—open-source software, a high school interaction study, and financial comentions—and find that hypergraph-based, theory-informed centrality better explains outcomes such as project success, student popularity, and same-day returns than standard graph centralities. For practice and policy, this yields more effective targeting, earlier warning signals, and improved allocation of attention and resources in collaborative work, public health, and market surveillance. We release an open-source Python package (HyperCentral) to support adoption.