Predicting Negative Ties in Social Networks

Abstract

Social networks play a pivotal role in connecting individuals and fostering interactions in various domains. They serve as platforms for communication, information sharing, and community building. Predicting the nature of relationships, specifically negative ties, within social networks has garnered significant attention due to its potential impact on user experiences and network dynamics. This study focuses on the prediction of negative ties in social networks, specifically in the context of the labeled Wikipedia Requests for Adminship online social network. Three distinct models are employed to accomplish this task. Firstly, a Light Gradient Boosting Model (LGBM) utilizes graph topology attributes to make predictions. The LGBM leverages the structural characteristics of the network, such as node centrality and connectivity, to identify negative ties. Secondly, a DistilBert language model is employed to process text data between users and their corresponding vote labels. The DistilBert model captures the semantic information embedded within the textual interactions, allowing for a more nuanced understanding of user sentiments and intentions. Finally, a Stacking Ensemble Model is employed to combine the predictions from the LGBM and DistilBert models. The Stacking Ensemble Model aggregates the predictions of the base models and employs a meta-learner to make the final predictions. Performance evaluation measures, including accuracy, precision, recall, F1-score, and elements of the confusion matrix, are used to assess the models’ predictive capabilities. Presently, all models exhibit strong performance in detecting positive and negative signed links within the network. Notably, the DistilBert and Stacking Ensemble models consistently demonstrate superior performance across all classes. Future research should focus on addressing class distribution issues, incorporating diverse data, and exploring ensemble techniques to further enhance the predictive capabilities of these models.