Fusion of Semantic, Visual and Network Information for Detection of Misinformation on Social Media

Fake news is the most prevailing buzzword in today's world. Fake news detection is a very time-consuming task that requires fact-checking either manually or automatically using machine learning techniques. The existing techniques focus only on the textual and image content. In this article, we have proposed a multimodal model-FakeMine for the mining of fake news content on social media. Our model explores the network structure of social media posts using Graph Neural Networks and combines them with semantic information from text and images in order to attain better accuracy. Our proposed work uses BERT for textual representations while preserving the semantic relationships in news articles. The image features are represented using VGG-19. The propagation structure of the circulating fake news is captured using Graph Neural Networks. All the features computed are fused together for a better classification. For classification, we have used LSTM which is optimized using Chimp Optimization. The FakeMine model was able to achieve an accuracy of 97.65 which exceeded all other baseline models for multiple modalities. It performed better than other models when tested on individual modalities as well. The proposed optimized LSTM classifier was also able to perform better than all other baseline classifiers.