Contemporary Machine Learning Methods For Twitter Bot Detection

Abstract

The aim of this thesis is to provide a comprehensive review of the current state of the art in detecting bots active on the social media platform of Twitter through the use of machine learning algorithms. Bots can be classified into a variety of categories and while some clearly identify themselves as automated accounts, many are posing as human users. This latter category is known to be used for a number of reasons like distorting online discourse and swaying political elections, manipulating stock markets, pushing conspiracy theories and spreading misinformation. Due to the ne- farious purposes these bots are used for it is important to evaluate current capabilities in detecting them and identify ways in which they can be improved. For this purpose four recently published research papers are examined, in which numerous machine learning methods are used on a common dataset, with varying results in accuracy and e ciency. These methods are reimplemented, in order to confirm the original authors’ findings, and in many cases enhanced by combining elements that proved e↵ective in other research e↵orts. A part of the work examined attempts to utilize the sentiment features of posted tweets, while other focuses on account and tweet level metadata. The rest of them follow a language-agnostic approach or try to classify ac- counts from single observations. All e↵orts above score highly in performance metrics (accuracy > 90%) and at least a couple of them achieve nearly perfect classification accuracy (AUC > 99%). Methods that use sentiment features demonstrate the need for better feature engineering in order to extract more features while the rest high- light the importance of further research in sampling techniques and the social aspect of making bot detection systems public and open source.