Harnessing interpretable novel combination of GloVe embedding with deep CNN-BiLSTM neural network for fake news detection

Abstract

The important issue of fake news to society is how it affects how society runs in terms of decision-making and public perception. Hence, this study is a comparative analysis of innovative hybrid deep learning models and embedding techniques focusing on interpretability using eXplainable Artificial Intelligence (XAI) for fake news detection. The popular fake news dataset is used to design and test a collection of state-of-the-art models, such as GloVe with CNN-BiLSTM, FastText-Bi-LSTM, and logistic regression with TF-IDF against the CNN and GloVe with BiLSTM and CNN models. In terms of accuracy, LSTM without FastText shows a performance of 98.33%, whereas GloVe with BiLSTM and CNN shows a 99.63% performance. Local Interpretable Model-Agnostic Explanations (LIME) is used to clarify how the input features make decisions on the high precision of the model. The integration of such state-of-the-art models with XAI is one of the major contributions of the study, which brings high accuracy as well as interpretability. Our study’s perspective addresses model performance and user trust in the future, laying the foundation for the practical implementation of reliable fake news detection systems.