Search and Justification of New Features for Enhancing Twitter Bot Detection's Accuracy Based on Genetic Algorithm and Naïve Bayes Approach

Twitter’s popularity has attracted both individuals and organizations. To be more efficient, the usage of bots has been adopted, which generate and share these tweets in an automated form. To this day, millions of accounts are known to be bots, which has interested researchers in locating ways in which bots can be identified on social media. The methods of identification vary in their implementation, algorithms, and chosen features. Therefore, the aim of this thesis is to locate new characteristics that can be used by algorithms to enhance social media bot detection on Twitter. The first part of the thesis concentrates on a literature research which provides already known features that help detecting bots. Using these features, a model based on Naïve Bayes algorithm is developed to classify Twitter users into human or bot accounts and its accuracy is tested. To find the combination of features a genetic algorithm is implemented on the same base. Lastly, new features are proposed and studied by examining how they affect the accuracy previously achieved. The research yielded 35 features commonly used for bot detection. Out of these, a combination of 15 achieved the highest accuracy when applying the classification algorithm. Introducing a new feature improves the accuracy, by using a different combination of 14 selected features. With the aim of discovering whether the achieved result is comparable with other tools, the algorithm’s score on a different dataset than before is then compared to “Botometer” where the later attained slightly poorer accuracy than the here presented classifier. In conclusion, a potentially new feature that enhances bot detection has been discovered which is defined by the linguistic complexity of a Twitter account. This feature can hence be further analysed and subsequently be implemented in already existing bot classification tools.