An Improved Method for Training Data Selection for Cross-Project Defect Prediction

The selection of relevant training data significantly improves the quality of cross-project defect prediction (CPDP) process. We propose a training data selection approach and compare its performance against the Burak filter and the Peter filter over Bug Prediction Dataset. In our approach (BurakMHD), firstly a data transformation is applied to the datasets. Then, individual instances of the target project adds k-instances at a minimum Hamming distance each from the transformed multi-source defective and non-defective data instances to the filtered training dataset (filtered TDS). Compared to using all the cross-project data, the false positive rate decreases by 10.6% associated with a 2.6% decrease in defect detection rate. The overall performance nMCC, Balance, G-measure increase by 2.9%, 5.7%, 6.6%, respectively. Compared to Burak filter and Peter filter, defect detection rate increases by 1.5% and 1.8%, respectively, and the false positive rate decreases by 6.4%. The overall performance nMCC, Balance, G-measure increase by 3%, 5.3%, 6.8% and by 3.2%, 5.5%, 7.1% compared to Burak and Peter filter, respectively. Compared to within-project predictions, the overall performance nMCC, Balance, G-measure increase by 1.1%, 3.4%, 4%, respectively, and the defect detection rate and false positive rate decrease by 9.2% and 13.1%, respectively. In general, our approach improved the performance significantly, compared to the Burak filter, Peter filter, cross-project prediction, and within-project prediction. Therefore, we conclude, applying data transformation and filtering training data separately from the defective and non-defective instances of cross-project data is helpful to select the relevant data for CPDP.