A Cost-Sensitive Shared Hidden Layer Autoencoder for Cross-Project Defect Prediction

摘要

Cross-project defect prediction means training a classifier model using the historical data of the other source project, and then testing whether the target project instance is defective or not. Since source and target projects have different data distributions, and data distribution difference will degrade the performance of classifier. Furthermore, the class imbalance of datasets increases the difficulty of classification. Therefore, a cost-sensitive shared hidden layer autoencoder (CSSHLA) method is proposed. CSSHLA learns a common feature representation between source and target projects by shared hidden layer autoencoder, and makes the different data distributions more similar. To solve the class imbalance problem, CSSHLA introduces a cost-sensitive factor to assign different importance weights to different instances. Experiments on 10 projects of PROMISE dataset show that CSSHLA improves the performance of cross-project defect prediction compared with baselines.