Predicting susceptibility to social bots on Twitter

The popularity of the Twitter social networking site has made it a target for social bots, which use increasingly-complex algorithms to engage users and pretend to be humans. While much research has studied how to identify such bots in the process of spam detection, little research has looked at the other side of the question - detecting users likely to be fooled by bots. In this paper, we examine a dataset consisting of 610 users who were messaged by Twitter bots, and determine which features describing these users were most helpful in predicting whether or not they would interact with the bots (through replies or following the bot). We then use six classifiers to build models for predicting whether a given user will interact with the bot, both using the selected features and using all features. We find that a users´ Klout score, friends count, and followers count are most predictive of whether a user will interact with a bot, and that the Random Forest algorithm produces the best classifier, when used in conjunction with one of the better feature ranking algorithms (although poor feature ranking can actually make performance worse than no feature ranking). Overall, these results show promise for helping understand which users are most vulnerable to social bots.