FakeIDCA: Fake news detection with incremental deep learning based concept drift adaption

Abstract

Social media facilitates rapid information sharing, improving exposure, connections, and content promotion. However, it also poses the challenge of fake news, which can mislead and harm individuals physically, and mentally, and incite violence. Fake news is often known as incorrect or misleading information. Prior research used Machine Learning (ML) and Deep Learning (DL) techniques for fake news detection. These studies predominantly relied on static offline models, overlooking the dynamic and evolving nature of news patterns, assuming their stability over time. Our paper proposes an incremental ensemble neural network for fake news detection that continuously learns from fake news streams, adapting to changes. It employs performance-based pruning to eliminate underperforming classifiers, improving overall performance. Additionally, the model detects concept drift in real-time and triggers adaptation strategies to maintain accuracy and robustness. The models undergo evaluation in two scenarios, utilizing consistent news patterns for training and testing, demonstrating consistent performance among all ML and incremental models. In the second scenario, the study analyzes the impact of news patterns over time, including concept drift due to significant events like the United States election. The analysis reveals that offline-trained methods are susceptible to performance degradation. However, the proposed model exhibits consistent performance with an accuracy of 97.90% and 99.76% on two fake news datasets, despite changes in the news pattern over time. The findings demonstrate how the evolution of the news pattern impacts the effectiveness of fake news detection models. The proposed model used for the experimentation indicates consistent performance even in the presence of drift