One-class classification approach for bot detection on Twitter

Twitter bots can be used for malicious purposes, such as distributing malware, disseminate fake news, faking grassroots political movements, and interfering with social movements.

Twitter allows its users to access its services through a web page, mobile applications, and an application programming interface (API). Even though API allows enhances users’ experience with contents, API also enables the creation of applications that control accounts and automate the posting of tweets.

Automation in Twitter can be performed in several ways. Fully automated accounts (bots), can be used for retweeting relevant content for specific communities or aggregate tweets about a topic. Partly automated accounts (cyborgs), allow users to automate specific tasks, such as scanning for mentions, schedule posts, and automatically respond to individual messages.

Detection of bots is an essential task for Twitter to suspend accounts that infringe on their terms and conditions. This detection uses mechanisms, such as allowing users to report or flag suspicious accounts. However, bots are getting more sophisticated to avoid detection.

In this study, Jorge Rodriguez-Ruiz and his colleagues used a one-class classification for bot detection on Twitter. One-class classifiers do not require examples of abnormal behaviour to discern between bots and humans. The authors compare the performance of their model to popular multi-class and one-class classifiers reported in the literature.

One-class classification is a kind of supervised classification, where one-class classifiers only require examples of the normal class to learn representations of it and identify if a new example belongs to that class or not. One-class classification is commonly used when the objective is finding deviations from normal or expected behaviour (anomaly detection).

Findings showed that one-class classifiers have a higher performance than other types of classifiers when detecting bot accounts with behaviours not present in the training set. However, classifiers designed to identify a specific type of bot had a higher performance than one-class classifiers.

The authors concluded that one-class classifiers could complement existing approaches for the detection of new bots.

To cite: Rodríguez-Ruiz, J., Mata-Sánchez. J. I., Monroy, R., Loyola-González, O. and López-Cuevas, A. (2020). A one-class classification approach for bot detection on Twitter. Computers & Security, 91.