Evaluating social bots detection approaches in different domains

Context: Social bots are automated users who make use of social networks to publish and interact with network users, mimicking or attempting to alter user behaviors with purposes such as spreading spam, malicious content, or misleading information, with various negative effects. Problem: Detecting these bots is a major challenge since, as detection mechanisms evolve, they are also enhanced to avoid such mechanisms, either by improving strategies for emulating real users or by organizing groups of bots in networks with the same purpose (botnets). IS Theory: The paper was developed considering Social Network Theory and Social Information Processing Theory. Method: The paper evaluates bots detection techniques by comparing the classifiers trained against three distinct datasets, aiming to emulate the behavior of a social network through time, to verify the performance of the classifiers in distinct conditions and the resilience of those techniques. Contributions and Impact in the IS area: The objective is to evaluate the effectiveness of the most common techniques in the domain in a variety of conditions based on the datasets used, an important challenge in the development and deployment of information systems. Summary of Results: The performance of the classifiers, when confronted against other datasets, was poor, showing that the classifiers trained for this purpose require constant maintenance to remain effective, reinforcing the need for improved techniques that are more resilient to changes over time and subject of messages. Proposed Solution: To counter those weaknesses, techniques that explore other characteristics, such as the message content, could be explored to improve the resilience of the classifiers.