Cost-Sensitive Hybrid Ensemble Deep Model for Software Defect Prediction on NASA Datasets

Software defect prediction is a crucial task for improving software reliability and reducing maintenance cost in large-scale software systems. One of the major challenges in defect prediction is severe class imbalance, where defective modules are significantly fewer than non-defective ones. Traditional machine learning models often fail to prioritize defect detection, leading to biased performance. This paper proposes a cost-sensitive hybrid ensemble deep model for software defect prediction using NASA benchmark datasets. The proposed framework integrates cost-sensitive learning with ensemble deep classification to enhance minority defect detection while maintaining overall predictive stability. A comprehensive pre-processing pipeline and multi-metric evaluation strategy are employed, including accuracy, precision, recall, F1-score, and ROC-AUC. Experimental results across eight NASA datasets demonstrate improved defective module detection and strong generalization capability. The proposed model provides a scalable and practical framework for intelligent software quality assurance in industrial environments.