Uncertainty-Aware Ensemble Models for Improved Defect Detection in Noisy Data

Abstract

Software defect prediction plays a crucial role in ensuring software quality and reliability, especially as modern systems become more complex and data rich. This study introduces an uncertainty-aware ensemble learning framework aimed at improving defect classification performance in noisy and imbalanced datasets, particularly those from the PROMISE and NASA KC1 repositories. The proposed model integrates multiple classifiers in a multi-learner ensemble structure to enhance generalization, improve true positive rates, and address the limitations of conventional single-model approaches. Key techniques include chi-square-based feature selection, ensemble pruning to avoid overfitting, and neural network-based classification through Extreme Learning Machines (ELMs). The methodology emphasizes the use of both homogeneous and heterogeneous ensembles, with training and prediction phases structured to handle data sparsity, high dimensionality, and class imbalance. Runtime experiments using decision trees, Naïve Bayes, and cost-sensitive learning demonstrated superior results for the ensemble model compared to traditional classifiers. Evaluation metrics such as accuracy, F-measure (0.9729), recall (0.7143), true positive rate (0.9857), and ROC AUC further validated the ensemble’s predictive robustness. Experimental results on the KC1 dataset showed that the proposed model outperformed baseline models in both accuracy and area under the ROC curve. Advanced data balancing techniques, including under-sampling, over-sampling, and active learning, were employed to improve the model’s ability to identify minority class instances. These findings suggest that uncertainty-aware ensemble approaches are effective tools for improving defect detection, particularly in noisy and imbalanced environments.