An explainable approach to fake news detection using PSO and mutual information

Fake news poses serious risks to societies by spreading disinformation rapidly and undermining trust in media. Automated methods for detecting fake news therefore need to be explainable to ensure transparency and trustworthiness. However, existing state-of-the-art methods are black-box and computationally expensive, making it difficult for practitioners and policymakers to understand and adopt them in real-world scenarios.

Feature selection can significantly enhance model transparency. Evolutionary computation (EC) methods are adaptable and naturally suited for optimization tasks such as feature selection. Among EC methods, Particle Swarm Optimization (PSO) requires few hyperparameters and converges quickly to good solutions. Therefore, this paper proposes an explainable approach to fake news classification using PSO. As mutual information (MI) quantifies the dependency between features and labels, the proposed method incorporates MI in its fitness function to guide the feature selection process and improve performance by eliminating redundant features.

Experimental results on three benchmark datasets (BuzzFeed, ISOT, and TruthSeeker) show that the proposed method achieved higher accuracy (85.21% on BuzzFeed, 99.99% on ISOT, and 90.92% on TruthSeeker) than baseline and state-of-the-art methods. Importantly, the proposed method selected a small feature subset, which helps explain model decisions, while requiring significantly less computation time.

These findings demonstrate the potential of PSO-based explainable approaches for practical and trustworthy fake news detection.