In view of common problems such as data imbalance and feature redundancy in software defect prediction

the sparse auto-encoder（ SAE） neural network is introduced and improved

a new classification model is proposed. Combining with advantages of SAE neural network and synthetic minority over-sampling technique（SMOTE）

the model can make up for the shortcomings of the traditional classification methods such as ignoring minority categories effects and not able to retain inner feature of data. The result of experiments base on some databases of NASA Metrics Data repository shows that the classification effect of proposed model is superior to other traditional algorithms in software defects prediction

especially improve the minority category classification precision of imbalanced datasets.