Combining clustering and classification algorithms for automatic bot detection - a case study on posts about COVID-19.

In the last decade, there has been a great insertion of bots in several social media. Among the potentially harmful effects of these software agents, there are: the spread of computer viruses and different internet scams, and the spread of fake news, with emphasis on political-electoral and public health-related news. This work presents a new approach for bots’ detection on Twitter, combining the use of feature selection, clustering, and classification algorithms. The proposed approach was compared with more conventional ones (for example, without the use of clustering) and the premise used in this work proved to be true: the use of clustering, together with the features selection, allowed the production of better classification models in order to identify not only the bots who have an activity profile considered non-human (extremely active on Twitter) but also other bots whose profiles are more similar to humans’ ones. The best results of automatic detection of bots reached an overall accuracy of 96.8% and F1 score equal to 0.622. As an additional advantage, these values were achieved by decision-tree models, which can be considered explainable artificial intelligence models.