Empirical evaluation of the active learning strategies on software defects prediction

摘要

Software defect prediction is a popular technical method in software engineering. In order to reduce the cost of a software defects, problems existing in the software are found by testing software products. Software defect prediction often uses machine learning techniques to improve the performance of software testing but requires enough labeled data when training the model. Because the cost of obtaining data is different from the label, the data is easy to obtain, but the label is cumbersome and expensive. In order to demonstrate software defect prediction, after the data obtained active learning algorithm is introduced to query the data, and the most valuable data is selected for expert annotation and then put into the model for training. However, it is not clear which active learning query strategy to choose the most effective in the software defect prediction model. We use different active learning strategy software defect prediction models for comparison. Experiment on the NASA dataset, using Naive Bayes and SVM, Linear Regression as the classifier. Comprehensive research results show that the Density-weighted strategy has a significant effect on the data set.