关键词: 软件缺陷预测, 深度学习, 类重叠, 语义特征

Abstract: Aiming at the problem that semantic features of source code were not fully used in the process of software defect prediction and class overlap in training data set, we proposed a cross-version software defect deep feature learning method for class overlap. This method used a hybrid nearest neighbor cleaning strategy to alleviate class overlap problem in deep learning semantic features. The test results on open data set PROMISE show that this strategy can improve the performance of software defect prediction based on deep semantic learning, and the classification performance can be improved by 14.8% at most in the median value. The experimental results show that a hybrid nearest neighbor cleaning strategy can be used to alleviate the class overlap problem in the cross-version deep defect prediction problem.

Key words: software defect prediction, deep learning, class overlap, semantic features