Abstract: Feature extraction is a key step in software defect prediction. The quality of the extracted features determines the performance of defect prediction. However, it is difficult for traditional feature extraction method to extract the deep nature features of software defect data. The auto-encoder model in the deep learning theory can automatically learn the features from original data and obtain its feature representation. Moreover, in order to enhance robustness of auto-encoder, a feature extraction method based on stacked denoising sparse auto-encoder is proposed. By setting different hidden layers, sparse parameter and noise increment methods, the required feature representation of classification and prediction is extracted directly and efficiently from software defect data. Experiment results using Eclipse defect dataset show that the proposed method has better prediction performance than traditional feature extraction method.

Key words: software defect prediction, feature extraction, deep learning, stacked denoising sparse auto-encoder