Cost-sensitive Approach for improving AUC-ROC Curve of Software Defect Prediction

Defects in software can cause various problems that are detrimental to many parties, from owners and developers to users. Predicting defects in software is one of the crucial factors that must be done before an application is launched. However, often when the prediction process is carried out, misclassification occurs, which causes losses at a later date. Misclassification usually occurs in datasets that are not balanced between majority and minority data. We suggest a combination of neural network (NN) and cost-sensitive method to minimize losses affected by misclassification. The author then compares the results of the evaluation of the proposed model with models without cost sensitivity to find out whether the model's performance has increased. Other classification algorithms such as decision tree (DT), logistic regression (LR), random forest (RF), and SVM will also be evaluated as part of a performance comparison of the proposed model based on AUC and ROC values in five different NASA datasets. The findings show that the cost-sensitive method can increase the AUC value in all datasets with an average increase of 13.6%. Cost-sensitive can overcome classification errors by increasing the AUC value in all models, and the optimal algorithm based on the average score in the proposed of neural network model.