Twitter Bot Detection with Reduced Feature Set

Abstract

Online social networks provide a novel channel to allow interaction between human beings. Its success has attracted interest in attacking and exploiting them through a wide range of unethical activities, such as malicious actions to manipulate users. One of the methods to carry out these abuses is the use of bots on Twitter. Recent examples of bots influencing public opinion in the election process demonstrate their potential harm to the democratic world. Such malicious behavior needs to be checked and its effects should be diminished. Recently, machine learning (ML) classifiers to distinguish between real and bot accounts have proven advances. Thus, in this work four ML algorithms were tested using a public dataset and a few expressive features based on simple user profile counters for the classification of bots on Twitter. We measured their performance compared to one state–of–the–art bot detection work. The classifier accuracy was considered homogeneous with a mean of 0.8549 and 0.1889 of standard deviation. Besides, all multiclass classifiers obtained AUCs greater than 0.9 indicating a practical benefit for bot detection on Twitter.