Combating Fake News on Social Media: A Fusion Approach for Improved Detection and Interpretability

The proliferation of fake news on social media prompted research groups to develop statistical and learning methods to combat this menace. Deep learning techniques could not model and improve in terms of adopting multi-transformer topologies, enhancing interpretability, and coping with uncertainty. This article suggests a fusion strategy to create a more reliable fake news detection (FND) model by fusing text and image features. The different combinations of information in single and multi-modalities have been investigated to find optimal conditions. In this paper, we have employed pre-trained models of Electra and XLnet for text feature learning. Furthermore, ELA has been used to highlight the modified image features and EfficientNetB0 for image learning. To enhance the interpretability of the proposed model, the superpixels contributing to its interpretability are identified using the Local Interpretable Model-agnostic Explanations (LIME). Three well-known datasets (Weibo, MediaEval, and CASIA) have been used in this study. The results show that employing ELA and LIME in conjunction with the fusion of text and image features provides a solid and understandable solution to the FND issue in social media compared to other techniques.