Word2Vec and LSTM based deep learning technique for context-free fake news detection

Nowadays, social media and virtual networking hubs like Twitter, Facebook have become an integral part of our daily lives. The recent boom in multimedia technology and increased internet access has led us into a hyper-connected global world. But these networks are often observed as the conduit for propagating fake news, which may cause a severe problem to a healthy social environment and destroy the harmony between the users. This calls for a proper segregation tool to classify various news articles as real or fake. Numerous research has been done on this topic, including the use of Artificial Intelligence (AI). In this work, we propose a deep learning based hybrid framework utilizing Word2Vec embedding and LSTM for fake news detection. As part of our approach, we generate Word2Vec embedding for obtaining vector representations of the news excerpts. The Word2Vec embeddings assist in generating context-free and data agnostic feature vectors for our news articles. The stacked LSTM layers process the extracted feature vectors to obtain the topic-relevant salient features for the news articles. This is followed by two fully connected dense layers for classifying whether the news excerpt under consideration is real or fake. We also perform hyperparameter tuning for achieving a better performance of our model. The proposed model is context-free and independent of datasets as well as topics for fake news detection. We compare the proposed method’s performance with some traditional Machine Learning baseline models, deep learning models, the pre-trained Bidirectional Encoder Representations from Transformers (BERT) via transfer learning, and some recently proposed state-of-the-art models. These models are tested on four datasets belonging to different domains for both training and testing purposes. Our proposed technique outperforms other well-known methods based on various performance metrics through intensive experimentation.