Analysis and predictability of centrality measures in competition networks
Anthony Bonato, Mariam Walaa
TL;DR
The paper tackles predicting node influence in directed competition networks by introducing the Common Out-Neighbor (CON) score and its dynamic, first-order and second-order variants. It formalizes CON via $\text{CON}_{1}$ and $\text{CON}_{2}$ based on adjacency matrices $\mathbf{A}$ and $\mathbf{A}_{2} = \mathbf{A} + \mathbf{A}^{2}$, and benchmarks it against traditional centralities. Using three real-world datasets—Survivor, Chess.com Titled Tuesday, and Dota 2—the authors build time-resolved networks, extract centrality features, and train supervised classifiers to predict ground-truth rankings; they report that CON often provides superior predictive power, with Dota 2 achieving the highest accuracy and Survivor the highest F1 score across models. They also analyze feature importance (MDI) and demonstrate strong correlations between CON-based rankings and ground truth, validating CON as a robust centerpiece for node ranking in adversarial networks. The work points to broad applicability to other competition-like systems and outlines avenues for extending CON to higher orders and to additional network models.
Abstract
The Common Out-Neighbor (or CON) score quantifies shared influence through outgoing links in competitive contexts. A dynamic analysis of competition networks reveals the CON score as a powerful predictor of node rankings. Defined in first-order and second-order forms, the CON score captures both direct and indirect competitive interactions, offering a comprehensive metric for evaluating node influence. Using datasets from Survivor, Chess.com, and Dota~2 online gaming competitions, directed competition networks are constructed, and the dynamic CON score is integrated into supervised machine learning models. Empirical results show that the CON score consistently outperforms traditional centrality measures such as PageRank, closeness, and betweenness centrality in classification tasks. By integrating dynamic centrality measures with machine learning, our proposed methodology accurately predicts outcomes in competition networks. The findings underline the CON score's robustness as a feature in node classification, offering a significant advancement in understanding and analyzing competitive interactions.