Towards an orthogonality constraint-based feature partitioning approach to classify veracity and identify stance overlapping of rumors on twitter

Abstract

The consequences of fake news and rumors have adversely affected social and political stability worldwide. Many such incidents have been reported, which resulted in mass chaos with the loss of lives and property. In response, many researchers have developed models for the veracity detection of rumors on social media. The recent advances in veracity detection models incorporate the use of Deep Neural Networks (DNN) over statistical and traditional machine learning based models. Current veracity detection approaches leverage powerful DNN models such as Transformer, Adversarial Networks, Graph Convolutional Networks (GCN), Variational Autoencoder (VAE) etc., along with exploiting the intuition of Multi-task learning (MTL) approach. In addition, most of these aforesaid well-known models rely on auxiliary (additional) information to a good extent. Presently, these recent models’ dependence on (1) auxiliary information and multiple tasks restrain productivity and incur cost on resources. Moreover, (2) the structural constraints of these models put a limiting effect on model deliverance. These two shortcomings of the recent models result in poor resource utilization and unstructured feature organization of the objective task, resulting in a compromised output of the model. In this paper, we present an efficient Segregated Non-overlapping and Collectively exhaustive DNN model (SeNoCe) which mitigates the effects of poor utilization of resources and enhances the model performance without the aid of auxiliary information or tasks which incur good manual efforts and costs. SeNoCe is capable of utilizing the fine-granularity of implicit features as attention for task identification. SeNoCe reports a major performance improvement over the state-of-the-art techniques on standard benchmark metrics across two real-world rumor datasets. It records a minimum of 24.06−12.8%, 53.1−49.2% improvement in terms of Macro F and Accuracy, respectively over the best performing state-of-the-art.