Ensembling Deep Learning Models for Fake News Classification

Abstract

Fake news remains one of the most pressing concerns of the current era for any citizen caring about political independence and the state of politics. In an implementation of NLP’s classic problems, we have set out to develop a stacked ensemble of deep learning models to classify and hence separate fake news from true articles. We have considered three well known deep learning architecture as baseline models, such as CNN, BiLSTM, and Multi Layer Perceptron (MLP). We also employed word embeddings like GloVe and Transformer BERT to boost our models’ performances, which brings our number of standalone models to six. We have further taken the individual model’s predictions to build a deep learning ensemble. We compared four such ensembles, namely models using hard weighted voting, soft voting, soft weighted voting, and logistic regression. After analysing the results, we conclude that our proposed logistic regression-based stacked ensemble is the best performing model, achieving significant f1-score of 96.03% and 74.24% using publicly available datasets.