ENHANCING MODEL GENERALIZATION IN CROSS PROJECT SOFTWARE DEFECT PREDICTION UNDER IMBALANCED DATA

Abstract

Strong machine learning frameworks that can detect problematic modules in a target project using data from many source projects, irrespective of data distribution or class balance, are necessary for cross-project software defect prediction (CPDP) using imbalanced data in order to improve model generalization. Over time, CPDP forecasts lose value due to numerous flawed cases. The majority is favored by several models. Data resampling (SMOTE and undersampling), cost-sensitive learning, transfer learning, domain adaptation, advanced techniques, and ensemble methods address this. It enhances model stability and minority class recognition. Distribution and sync alignment promote cross-field knowledge and minimize project discrepancies. This uses imbalance control and generalization to improve memory, F1-score, and AUC. Fault-prediction systems become more reliable and scalable in many software development scenarios.