Concept Drift Influenced by Topic Change in Data Streams from Social Media
Martin Sarnovský, František Babič · 2025
The necessity for detecting antisocial behaviour on the web is one of the most talked about and important topics in recent years. This kind of behaviour has of multiple forms, for example hate speech in discussions, fake reviews on the on-line stores, or spreading of the fake news. The focus of this work is aimed on the detection of fake news and misinformation in online posts on social media platforms. Many studies have focused on the topic of fake news detection before, but did not take into consideration that the nature of online posts is everchanging and dynamic. Topics in online discussion change and evolve over time, as well as topics of misinformation and fake news. Such a change may cause the concept drift, a change in learnt concepts, which may influence the performance of the detection models. To confirm this assumption, we create a synthetic data stream of multi topic social media posts with simulated topic change. Our main objective was to confirm whether such a change results in concept drift and how it affects the predictive performance of the standard models in comparison to adaptive classifiers.