CAWE-X: A Confidence-Aware Explainable Weighted Ensemble Framework for Misleading Online Content Detection with Calibrated Uncertainty

Abstract

Detection of potentially deceptive content from the web is generally considered as the choice of one optimal classifier, but in practice, several classifiers based on the same representation produce nearly equal performance while having different classifications on particular objects, have poor prediction confidence and cannot provide explanations for their decisions. This work introduces the CAWE-X model which is the Confidence-Aware Explainable Weighted Ensemble approach which combines four complementary models – Multinomial Naive Bayes, Logistic Regression, Linear Support Vector Machine and Random Forest using TF-IDF n-gram representation – into one collaborative classification mechanism. Calibration of the scores of the learners is performed using either Platt scaling or isotonic regression based on the out-of-fold predictions, while weighting instances is a combination of validation skills of the learners and their prediction confidence. The framework outputs a Final Misleading Score, abstains with a Requires Verification label when the fused confidence is low or the learners are evenly split, and explains every decision through ensemble-weighted n-gram attributions and per-learner contribution shares. The framework is evaluated on LIAR and FakeNewsNet (PolitiFact and GossipCop) with leakage control, staged ablation, cross-dataset validation, a DistilBERT baseline, bootstrap confidence intervals, statistical tests, calibration analysis, risk–coverage analysis, explanation faithfulness testing and structured error analysis. In the illustrative LIAR results, CAWE-X achieved a macro-F1 of 0.683 and ROC-AUC of 0.756, compared with 0.660/0.729 for the strongest single conventional learner (SVM) and 0.692/0.771 for DistilBERT. Calibration reduced ensemble ECE from 0.049 to 0.026. At the selected abstention threshold (τ = 0.10), coverage was 0.801 and selective macro-F1 increased to 0.733. Cross-dataset evaluation showed the expected degradation under distribution shift, while CAWE-X retained ROC-AUC values of 0.667 for LIAR → GossipCop and 0.713 for LIAR → PolitiFact.