Enhancing Agile Defect Prediction with Optimized Machine Learning and Feature Selection

In Agile software development, efficient defect prediction is crucial because of the rapid and iterative nature of the delivery. Conventional methods that rely on source code or commit logs often fail to capture the critical contextual signals necessary for early bug detection. This study proposes a hybrid machine learning framework that leverages enriched contextual features from Jira issue tickets and combines them with optimized feature selection techniques. Various classification models, including Random Forest, XGBoost, CatBoost, SVM, and Transformer, are employed to predict defects. To further enhance model performance, metaheuristic-based feature selection methods such as the Bat Algorithm (BA) and Particle Swarm Optimization (PSO) are applied to reduce dimensionality and improve predictive relevance. Experimental results show that Random Forest with BA optimization achieves the highest performance, with an F1-score of 0.83 and an AUC-ROC of 0.86, outperforming other models. While the Transformer model does not surpass tree-based algorithms in all metrics, it shows high recall and competitive F1-scores, making it suitable for high-sensitivity applications. These findings highlight the importance of integrating optimized machine learning models and feature selection techniques to improve model robustness, reduce computational complexity, and meet the needs of Agile development. This approach supports software teams in prioritizing quality assurance tasks, reducing long-term maintenance costs, and optimizing defect management processes.