Isolation Forest Filter to Simplify Training Data for Cross-Project Defect Prediction

Cross-project defect prediction (CPDP) is an active research area. When the historical data is limited or a new project is to develop, establishing CPDP models is very useful, which assists software testers to judge the defect-prone entities or software managers to focus on the "important" parts by allocating the manpower, budget, time. However, the dissimilarity of data distributions between the source projects and the target project decreases the performance of CPDP models. How to simplify or the cross-project training data is an important problem. To solve this issue, an isolation forest (iForest) filter is proposed. We use 15 versions of different java projects from open PROMISE Data Repository and five typical predictors (naïve bayes (NB), decision tree (DT), logistic regression(LR), k-nearest neighbor(k-NN) and random forest(RF) to build 1050 (15*14*5) software defect prediction models (SDPM). Meanwhile, we compare our models with Burak Filter models and Peter Filter models. From the results of performance measures, called AUC, balance, G-measure, G-mean, F1-measure, we can know that our iForest filter is feasible and even better than other two. Therefore, using iForest filter can make cross-project training data simple and build efficient SDPM.