MULTI-VIEW USER SIMILARITY ANALYSIS ON TWITTER

Abstract

With the rise of social media sites like Twitter, there is a huge amount of user-generated data available, making possible various applications like recommendations, community finding, user profiling, and targeted marketing. User similarity on Twitter can be a difficult problem to solve because social media data consists of avariety ofcomponents, including text data, behaviors, user profiles, and connections. The traditional method of solving the user similarity problem by using only one type of data does not work effectively anymore. This thesis offers a systematic study and comparative analysis of user similarity techniques on Twitter. The review covers previous research in this area, classifies techniques based on the type of their work (content-based, graph-based, signal-based, or embedding-based), and discusses main difficulties of this research such as data sparsity, noise in text, scalability, and lack of benchmarks. It also notes the emerging trend to move away from feature-based techniques towards hybrid and representation learning-based techniques. The experimental comparison study compares three types of similarity models: those based on features, representation learning, and metadata. The feature-based method is based on semantic, behavioral, and structural similarity metrics, while the representation learning method relies on the use of Sentence-BERT embeddings and temporal behavioral modeling to determine latent relationships between users. Metadata-based similarity approach depends on profile information, textual similarity, and network-based methods. From theexperimentsconducted, itcanbeseenthatthesimilaritiesbetweenusersin social media sites are multi-dimensional in nature and therefore cannot be captured by just one methodalone. Thefeature-basedapproachismoreinterpretable,representation learning offers better latent connections, and the metadata-driven approach performs better in terms of classification. In conclusion, it can be said that hybrid and multi-view approaches for measuring similarity are more effective.