USER SIMILARITY ON TWITTER: A DUAL PERSPECTIVE OF LITERATURE REVIEW AND EXPERIMENTAL COMPARISON

Abstract

The rapid proliferation of social media platforms, particularly Twitter, has necessitated advanced techniques for understanding user behavior, identifying similar users, and uncovering community structures. This thesis presents a comprehensive study of methods for detecting user similarity and communities on Twitter, encompassing both literature review and comparative analysis perspectives. The first part of the research synthesizes existing approaches into three primary categories: signal-based, machine-learning-based, and graph-based methods. These approaches leverage interaction patterns, social graph structures, and content alignment to address applications in security, audience targeting, and social recommendation. The strengths, limitations, and scalability of these methods are critically evaluated, with an emphasis on their adaptability to real-world scenarios and societal implications. The second part focuses on a detailed comparative analysis of three established user similarity frameworks: TSim, Characterizing and Detecting Similar Twitter Users, and Self-Similarity of Twitter Users. Utilizing a dataset derived from the Twitter API, the study implemented ten similarity signals encompassing interaction, content, and network-based metrics. The results highlight strong correlations between interaction and retweet similarity metrics while underscoring the complementary insights of profile-based features. The computed rankings, derived from an aggregated similarity score, achieved a high Spearman correlation of 0.91 with human evaluations, validating the model's effectiveness. This thesis concludes by identifying limitations and proposing future directions, such as adaptive weighting strategies, integration of temporal dynamics, and scalability testing for large datasets. By bridging theoretical insights with practical applications, this work contributes to the development of robust, adaptive, and interpretable systems for similarity detection and community discovery, enhancing the personalization and utility of social media platforms.