Toward a new approach to author profiling based on the extraction of statistical features

• Published in: Social network analysis and mining Vol. 11; no. 1; p. 59
• Main Authors: Ouni, Sarra, Fkih, Fethi, Omri, Mohamed Nazih
• Format: Journal Article
• Language: English
• Published: Vienna Springer Vienna 01.12.2021; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Applications of Graph Theory and Complex Networks; Classification; Computer Science; Cybernetics; Data Mining and Knowledge Discovery; Datasets; Economics; Extraction; False information; Forensic medicine; Game Theory; Gender; Humanities; Law; Machine learning; Marketing; Methodology of the Social Sciences; Original Article; Performance evaluation; Perpetrators; Privacy; Profiles; Social and Behav. Sciences; Social media; Social networks; Software agents; Statistics for Social Sciences; User behavior; Bot detection; Twitter; Gender detection; Features extraction; Machine learning
• ISSN: 1869-5450; 1869-5469
• DOI: 10.1007/s13278-021-00768-6

Recently, author profiling on social media and on online platforms, characterized by a huge volumes of data, has become more than a critical issue. This issue is of increasing interest in various fields related to forensic medicine, security, marketing, education, etc. The main objective of author profiling is to identify the type of writer of the messages, whether it is a human or a bot with a very strong presence. These bots have the task of drawing the attention of browsers to specific events, often used to disseminate incorrect and/or false information. In this work, we offer a new approach to detect these bots and the kind of anonymous perpetrators on these social networks. Our approach, purely statistical, is based on digital features (APSF), extracted from users’ tweets, and on the technique of random forests. A total of 17 stylometry-based features were used to train the model. To assess the performance of our approach, we considered different standard measures, namely accuracy, precision, recall and F1-score. The results obtained show that our approach gives the best performance for both English and Spanish languages. For the English dataset, we achieved an accuracy of 92.45% for the bot detection task and 90.36% for the gender classification; similarly, we obtained accuracy values of 89.68% and 88.88% for the Spanish dataset.