AEC: A novel adaptive ensemble classifier with LIME and SHAP-Based interpretability for fake news detection

The constant availability of fake news on social media and other information-sharing platforms has raised a demand for efficient and effective fake news detection models. Current solutions propose the principle of static feature extraction, the dependence on a single classifier, or modest results that do not meet the requirements in today’s environment with its complexity of data. As a result of these, this study presents the Adaptive Ensemble Classifier (AEC), a new ensemble system that consists of hybrid decision trees and Support Vector Machine (SVM) similar to margin optimization for the improvement of classification performance. The proposed AEC incorporates several innovative features: dynamic feature selection through adaptive neighbourhood selection based on feature importance, SVM-based refinement of decision boundaries for improved precision, and a weighted ensemble voting mechanism to ensure robust predictions. In addition, to ensuring explain ability, the system uses LIME and SHAP to provide probability-based explanations for the predictions and the features that influence the results. The performance of the AEC is evaluated using public datasets such as the Fake News dataset and cross-domain performance using COVID-19 Fake News Dataset. Experimental results confirmed that the proposed model achieved an impressive accuracy of 99.74% compared to traditional Machine Learning (ML) and Deep Learning (DL) models in particular in aspects of accuracy, precision, recall, and F1 score. The computational efficiency is also evaluated by comparing training time, memory usage, peak memory usage, inference time, and model size with existing models. The interpretability offered by LIME focuses on the most important features that affect predictions, which makes the system more useful in real-world situations. Finally, the different statistical analysis tests are also employed on proposed AEC and existing ML and DL models such as Paired t-test, Kruskal-Wallis, Dunn’s, Bootstrap Resampling, Cohen’s d Effect Size and Confidence Intervals. The results showed that significant performance difference between proposed AEC and other models.