A Hybrid Model for the Measurement of the Similarity between Twitter Profiles

Abstract

Social media platforms have been an undeniable part of our lifestyle for the past decade. Analyzing the information that is being shared is a crucial step to understanding human behavior. Social media analysis aims to guarantee a better experience for the user and to increase user satisfaction. To draw any further conclusions, first, it is necessary to know how to compare users. In this paper, a hybrid model is proposed to measure the degree of similarity between Twitter profiles by calculating features related to the users’ behavioral habits. For this, first, the timeline of each profile was extracted using the official TwitterAPI. Then, three aspects of a profile were deliberated in parallel. Behavioral ratios are time-series-related information showing the consistency and habits of the user. Dynamic time warping was utilized to compare the behavioral ratios of two profiles. Next, the audience network was extracted for each user, and to estimate the similarity of two sets, the Jaccard similarity was used. Finally, for the content similarity measurement, the tweets were preprocessed using the feature extraction method; TF-IDF and DistilBERT were employed for feature extraction and then compared using the cosine similarity method. The results showed that TF-IDF had slightly better performance; it was therefore selected for use in the model. When measuring the similarity level of different profiles, a Random Forest classification model was used, which was trained on 19,900 users, revealing a 0.97 accuracy in detecting similar profiles from different ones. As a step further, this convoluted similarity measurement can find users with very short distances, which are indicative of duplicate users.