Statistics of the dataset.

The proliferation of fake news is one of the major problems that causes personal and societal harm. In today’s fast-paced digital age, misinformation spreads rapidly, often leaving individuals without the time to verify the authenticity of the information. This can cause irreparable damage to personal reputations and organizational credibility. Thus, instigated by the quintessential necessity, there is a dire need to construct a framework for the automatic detection and identification of fake news at its inception. This research presents a novel approach that leverages a combination of three popular word embeddings (FastText, FastText-Subword, and GloVe) integrated with a customized convolutional neural-network(CNN) to classify fake news accurately. The proposed model was tested against the Fake News Challenge dataset. Hundreds of word vector features were generated from the combined embedding and then managed with PCA and significant features were extracted. The proposed model gives an accuracy of 94.58%, 95.35% precision, 97.29% recall, and an F1 score of 96.11%. The proposed framework’s robustness is demonstrated when compared with other machine, deep, and ensemble learning approaches, showing superior performance. Furthermore, the effectiveness of the model is validated on an independent Arabic Fake News dataset.