From Political Encyclopedias to Hypernetworks: A Case Study of Hate Groups and Conspiracy Theories

We analyze the contents of two encyclopedias: (1) Barry J. Balleck’s "Hate Groups and Extremist Organizations in America" (ABC-CLIO, 2019) and (2) Peter Knight (ed.), "Conspiracy Theories in American History" (ABC-CLIO, 2003). Using standard NLP techniques, we construct directed graphs for each encyclopedia, where nodes represent individual entries and edges denote references to other entries within the text of each entry. For the entries in each encyclopedia, represented as nodes in a directed graph, we compute various centrality indices, including out-degree, in-degree, closeness, betweenness, eigenvector, HITS (hubs and authorities), Katz, PageRank, and load centrality. Additionally, we apply two community detection algorithms—Girvan-Newman and asynchronous label propagation—to group these entries into clusters, or "communities." We further model these directed graphs as hypergraphs (hypernetworks), where hyperedges are defined for each entry as the set of all referenced entries. Specifically, we present results of hyperlink prediction on these hypergraphs, a natural extension of link prediction in graphs, which aims to infer missing hyperlinks in hypergraphs. Hyperlink prediction has diverse applications in fields such as metabolic networks, protein-protein interaction networks, chemical reaction networks, bibliometric networks, and social communication networks. Our work represents one of the first attempts to address challenges in political networks using hypergraph modeling and hyperlink prediction. In this study, we systematically demonstrate hyperlink prediction for the entries in two encyclopedias, leveraging Python and the PyTorch library. We explore two hyperlink prediction techniques: a deep learning-based method (Chebyshev Spectral Hyperlink Predictor, or CHESHIRE) and a structural similarity-based method (Hyperlink Prediction Using Resource Allocation, or HPRA). To evaluate the performance of these methods, we employ several metrics commonly used in machine learning classification tasks, including F1 score, ROC AUC, accuracy, precision, recall, log loss, and Matthews correlation coefficient. The motivation for applying hyperlink prediction to political networks derived from encyclopedia texts lies in its potential to uncover relational consistency among entries. We approach the problem of hypergraph reconstruction as a subtask of hyperlink prediction by hiding a subset of entries and predicting the likelihood of reconstructing the hidden hyperlinks based on the structural characteristics of the observed entries. By applying hyperlink prediction to encyclopedias, we aim to systematically analyze and better understand the structural coherence of encyclopedic entries and their links (references to other entries).