• Nilakantha Sahu
• A Pratik Kumar
• Rounak Mahto
• Hrushikesh pathy
• Santosh Kumar Kar

Software defect forecasting is vital in enhancing software reliability and helping to minimize the cost of software maintenance in that defect-prone software components can be detected at an early stage of the software development life cycle. The foundation paper examined various machine learning and deep learning models based on the Unified Bug Dataset, the database of software metrics applied to various open-source projects. A deep learning Long Short-Term Memory (LSTM) model has demonstrated the highest performance with an approximate accuracy of 0.87, yet the research was restricted by a lack of optimization of the decision threshold, imbalance in classes, and redundant software metrics. To enable improvement of the situation, this study suggests a superior software defect prediction model incorporating advanced preprocessing, and ensemble learning models. The methodology suggested will involve preprocessing of features by using the standardization technique, the mutual information-based feature selection algorithm to eliminate redundancy measures, and Borderline-SMOTE to balance out classes. In addition, stacking ensemble model, which incorporates Random Forest, XGBoost and Multilayer Perceptron (MLP) is constructed in order to take advantage of multiple learners. The meta-classifier used is logistic regression to enhance the strength of the prediction. Moreover, decision-threshold optimization is used to optimize the Binary F1-score, which is especially relevant to identifying defect-prone modules on imbalanced datasets. Experimental results on the Unified Bug Dataset show that the given framework significantly enhances the results of defect prediction. The ensemble stacking obtained the accuracy of 94.90, Macro F1-score 0.799, Weighted F1-score 0.948 and Binary F1-score 0.626 that were higher compared to the baseline deep learning model (LSTM accuracy approx. to 0.87). Other metrics where the model performed well are ROC-AUC (0.954), PR-AUC (0.690) and Balanced Accuracy (0.787) and Cohens Kappa (0.598). These findings suggest that ensemble learning, imbalance and threshold optimization are better predictors of software defects in large scale software systems, as a combination.

Software Defect Prediction, Machine Learning, Ensemble Learning, Feature Selection, Imbalanced Data Handling.

"An Optimized Machine Learning Framework for Software Defect Prediction Using the Unified Bug Dataset", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.13, Issue 3, page no.c639-c648, March-2026, Available :http://www.jetir.org/papers/JETIR2603283.pdf

"An Optimized Machine Learning Framework for Software Defect Prediction Using the Unified Bug Dataset", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.13, Issue 3, page no. ppc639-c648, March-2026, Available at : http://www.jetir.org/papers/JETIR2603283.pdf