Predicting Voting Behavior in Online Social Platforms Using Machine Learning

Abstract

Modeling voting behavior in online social networks is challenging because of the visibility of prior votes and the ability to view and respond to others' voting choices. This paper uses the Wikipedia Request for Adminship (RFA), in which the task is formulated as a binary classification problem, where voters cast ballots either supporting or opposing a nominated administrator candidate. Features tested include behavioral history, sentiment derived from comments, graph-based characteristics, and temporal herding. Three machine learning algorithms were tested, namely Logistic Regression, Random Forest, and XGBoost. Behavioral features alone provide limited predictive performance. Adding sentiment features did not lead to a noticeable improvement. However, Adding sentiment and graph-based features resulted in only marginal improvements.The top-performing model using the full feature set was XGBoost, achieving an Accuracy of 84.01%, a Macro F1-Score of 77.24%, and a Precision-Recall Area Under Curve (PR-AUC) of 95.82%. These results show that overall, voters make their voting decisions based on previous voting decisions from other voters as well as the characteristics of the specific voter making that decision.