Analysis of social networks content to identify fake news using stacked combination of deep neural networks

Abstract

In today’s fast-paced world, the unprecedented expansion of social networks and the huge volume of information has made automatic detection of fake news an undeniable necessity. The dissemination of fake news and misinformation can have a devastating impact on public opinion and social decision-making. This challenge requires new and powerful approaches in the fields of deep learning and natural language processing to accurately and quickly identify fake news and prevent its dissemination. For that purpose, this current work presents a new and efficient solution to detecting and spotting spurious news on social media. This method, through deep text content analysis and the employment of advanced deep learning techniques, aims to provide an expansive and accurate response to solve this problem. The proposed method consists of three determining steps: 1) The input data is initially prepared for the next steps using preprocessing techniques. This is done through noise removal, text normalization, and data conversion into a format that can be processed by deep learning models. 2) A hybrid method is then used to extract text features, which is a combination of a list of statistical features (e.g., text length, word count, and links), GloVe-based semantic features (to represent the word relationships), and Character N-Grams (CNG) (to improve misspelling and linguistic anomaly robustness). 3) Finally, for each set of features, a particular deep model is trained to predict based on each component. Specifically, a Multilayer Perceptron (MLP) model is used for statistical feature analysis, and Convolutional Neural Network (CNN) models are used for GloVe and CNG features. Both models generate individual predictions from the input features presented to them, and the predicted labels and the posterior probability vector for each of the models are combined to output a vector to be forwarded to the meta-learner (a MLP model). By learning patterns in the combinations of outputs and the probability vectors of the individual base models, the MLP model can correctly identify fake news or real news. Experimental results conducted on two authentic datasets, GossipCop and Politifact, show that our proposed method achieves 99.45 % and 97.40 % accuracies, respectively. This achievement indicates the very good and effective performance of our method in detecting fake news on both datasets.