The Impact of Data Splitting Techniques on the Accuracy of Rumour Detection Models in Microblogging Platforms: A Comparative Analysis

The swift propagation of online misinformation present a significant challenge, primarily due to the lack of supervision to control it. Automated false information detection methods are crucial to mitigate the adverse effect of false information. Conventional techniques have depended heavily on manual text feature extraction that is both time-consuming and less effective. To overcome this issue, rather than relying on traditional feature extraction methods, we propose a new technique by leveraging sentence embeddings produced by different transformer models. Transformer models excel in capturing the contextual meaning of words, allowing us to represent each post as a vector based on its context. These vectors are then classified as rumors or non-rumors using different supervised learning techniques. Our work offers valuable contributions in two key domains: rumor detection on ‘X’ platform and the evaluation of text embeddings generated by various transformer models, with a focus on how different data splitting techniques influence model performance. By testing different splitting strategies, this study explores how these methods affect the models' ability to generalize and detect rumors accurately. The findings will shed light on the connection between data splitting techniques and model effectiveness, providing guidance for developing strong rumor detection systems in social media environments.