Eliminating the high false?positive rate in defect prediction through <scp>BayesNet</scp> with adjustable weight

In defect prediction, a high false-positive rate (FPR) caused by class imbalance not only increases the workload of testing and development but also consumes unnecessary costs. Many models against have been proposed to improve accuracy their ability reduce FPR is unclear. To solve these problems, we first BayesNet with adjustable weights, called WBN, in software which an algorithm independent data preprocessing techniques. The mechanism our WBN change sampling probability misclassified instances when training model, making model focus more on false alarm instances. And then, investigate five mainstream for solving select them as comparison test validity methods. experimental result eight open-source projects shows that a) in-version prediction (IVDP) cross-version (CVDP), effectively reduces means 0.384 0.322, respectively; b) compared improved subclass discriminant analysis (ISDA) lowest all control models, reduced maintained recall whose mean value was 0.797, whereas ISDA did not, average 0.397; c) CVDP, FPR, has significant superiority over baseline. Besides, found difference between set impact CVDP performance, recommending practitioners choose best dataset from historical version through special technology.