Polarity-Aware Fake News Detection using Machine Learning Algorithms

Abstract

The tremendous evolution and widespread use of online and social media platforms have led to the easy publication of various types of fake news that pose considerable threats to social stability, public confidence, and informed decision-making. Many machine learning (ML) and deep learning (DL) techniques have been employed to solve this problem in recent studies, but often suffer from high computational cost, their dependency on relatively small and domain-specific datasets, or even not reaching the desired evaluation outcomes. In this paper, we propose a robust machine learning-based fake news detector that integrates linguistic, structural, sentiment, and phrase-level features to enhance classification performance while maintaining computational efficiency. To train the model, a large-scale balanced dataset was constructed by combining and strictly filtering multiple open-source benchmarks, resulting in a clean and revised dataset consisting of about 130k English news articles. The extracted feature set combines lexical statistics, part-of-speech (POS) distributions, n-gram phrase patterns, and sentiment polarity indications. Several supervised machine learning algorithms were trained on the same dataset and evaluated under identical experimental situations. The experimental results prove that the XGBoost-based model significantly beats all other baselines, attaining an accuracy of 99.37%, an area under the curve (AUC) of 0.9998, and exhibiting extremely stable cross-validation performance. The analysis of feature importance further proves that phrase-level and overall sentiment polarity are the most powerful predictors. These results show that XGBoost with carefully extracted linguistic features can exceed preceding deep learning–based methods, offering an effective, clear, and scalable solution for the real-world fake news problem.