A temporal ensembling based semi-supervised ConvNet for the detection of fake news articles

摘要

Internet-based information circulation has given rise to the proliferation of fake and misleading contents, which has extreme hostile effects on individuals and humanity. Supervised artificial intelligence techniques require a huge amount of annotated data which is a time-consuming, expensive and laborious task as the speed and volume of social media news generation is very high. To counter this situation, we propose an innovative Convolutional Neural Network semi-supervised framework built on the self-ensembling concept to take leverage of the linguistic and stylometric information of annotated news articles, at the same time explore the hidden patterns in unlabelled data as well. Self-ensembling provides consensus predictions of the labels of unannotated data using previous epochs outputs of network-in-training. These accumulated ensemble predictions are supposed to be a better predictor for the unknown labels than the output of most recent training epoch, thus suitable to be used as a proxy for the labels of unannotated data. The uniqueness of the framework is that it ensembles all the outputs of previous training epochs of the neural network to use them as an unsupervised target for comparing them with current output prediction of unlabelled articles. The framework is validated with extensive experiments on three datasets for different proportions of labelled and unlabelled data. It can achieve highest 97.45% fake news classification accuracy using 50% labelled articles on Fake News Data Kaggle dataset. Contemporary baseline methods are placed in juxtaposition with the proposed architecture which demonstrates the robustness of our work compared to the state-of-the-art.