Improvement of a Machine Learning Model Using a Sentiment Analysis Algorithm to Detect Fake News

Abstract

These days, the problem of fake news has grown to be a major social and personal concern. With the amount of information generated through social media, it is very crucial to be able to detect and properly take care of that fake information. Previous studies proposed a machine learning model to detect fake news in online Thai health and medical articles. Still, the problem of detecting fake news with similar content but different objectives exists, and the accuracy of the model needs improvement. Therefore, this study aims to solve these problems by adding 33 new features, including textual features, sentiment-based features, and lexicon features, i.e., herbs, fruits, and vegetables, to identify the objective of an article. We trained and tested the model's prediction accuracy on a new dataset containing 582 reliable and 435 unreliable (fake news) articles from eight Thai websites. Our improved classification model using XGBoost with Lasso, the best feature selection method, achieved an accuracy of 97.76% without over-fitting, reflecting a 7.16% improvement over our earlier model.