Fake News Detection Using Lightweight Machine Learning Models

Abstract

Justice and public trust have been harmed as a result of the massive spread of false news and the deterioration of democracy. Fake news detection is crucial for maintaining a healthy and trustworthy news ecosystem, which is why it is a growing global concern. Despite the fact that there are various methodologies for detecting fake news, style-based fake news detection algorithms consistently outperform the other models by overcoming the disadvantage of recognizing fake news before it spreads. It checks the validity of news by extracting language elements at four different levels. They do, however, provide a number of obstacles, such as the growing linguistic style of fake news and the use of a variety of linguistic words. As a result, a thorough examination of style-based features and models is required. In this research, we investigate the identification of false news using lexical features collected from three separate datasets and various lightweight machine learning models. SVM, logistic regression, decision tree, and neural networks are the models we used to conduct the case study. We investigate each model by plotting ROC curves for each dataset. We observed that SVM and logistic regression substantially rely on quantity or quality lexical features, i.e., more textual content of the training examples, based on their data gathered from the ROC curve. The size of the dataset is more important to the decision tree. However, when comparing their overall performance on, it can be concluded that logistic regression is more efficient for classification than the others, with an average performance of 90%.