Cross-Project Defect Prediction Model Based on Enhanced Transfer Learning

This study proposes a two-stage Cross-Project Defect Prediction (CPDP) framework to address class and feature distribution imbalance. It uses adversarial training-enhanced transfer learning to identify defective software modules. Its superiority is shown through experiments on public datasets and comparison with traditional models. The first stage involves feature extraction with two encoders and advanced preprocessing techniques. The second stage utilizes transfer learning, ensemble learning, fine-tuning, and the Synthetic Minority Oversampling Technique (SMOTE) method. Future research can expand its application and optimize the model to handle complex imbalances or incorporate other techniques to enhance predictive performance, increasing the practical potential of the CPDP model in software engineering.