| With the advancement of social information construction,the quantity of software increases exponentially,and the quality of software becomes the focus of people’s attention.Therefore,the software defect prediction technology is used to predict the modules that may have defects in software,so as to ensure the software quality,which has become the research focus in the software field.At present,in the process of building software defect prediction model,there are three main problems: decision surface offset,integrated model overfitting and the same cost of misclassified samples.To solve these problems,this paper conducts the following research on the level of data sampling and integration algorithm:(1)Data Sampling level: In order to solve the problem that the offset of decision surface and the cost of misclassified samples are the same,a fusion cost sensitive error sampling algorithm(Cost-sensitive Adapt Misclassification Sampling & SMOTE,CSAMS&S)is proposed.First,The algorithm uses naive Bayes classifier to train and predict the original data set to get the misclassified data set and the correct classified data set.Then,in order to solve the sparsity of sample in the center area of defect class,use SMOTE method to sample the correct classified data set.In order to solve the problem of the same cost of misclassification,the CS-AMS algorithm was used to sample misclassification data set.The CS-AMS algorithm innovatively fuses the cost factor into the sampling field,and adaptively synthesizes "artificial" samples from the two types of samples at the boundary,thus achieving the local equilibrium state of the boundary region.A new interpolation strategy based on the class of nearest neighbor samples is proposed when synthesizing artificial samples.Comparative experiments were conducted in AEEEM,NASA and Relink datasets.With KNN,decision tree and random forest algorithm as classifiers and F1 and G-Mean as evaluation indexes,the results show that CS-AMS&S algorithm has good performance.(2)Classification level of integration algorithm: In order to improve the classification performance and generalization ability of the model,theoretical analysis of LightGBM integration algorithm(Light Gradient Boosting Machine)is carried out,and the core parameters of its grid search method is used to obtain the optimal parameters of the corresponding data set.It is proved that LightGBM is an integrated classification algorithm with high performance by comparing with various classifiers.(3)Model construction stage: Based on the above research content,a cost-sensitive misclassification sampling integrated software defect prediction model(CS-AMS&S LightGBM)is constructed.Firstly,the feasibility of the combination of sampling and LightGBM is proved.Then,the prediction model of CS-AMS&S combined with LightGBM is compared with a variety of cost sensitive and sampling integrated prediction models.The experimental results show that this model has good accuracy and classification performance in the field of software defect prediction,and is an efficient software defect prediction model. |
