REPD: Source code defect prediction as anomaly detection

Abstract

In this paper, we present a novel approach for within-project source code defect prediction. Since defect prediction datasets are typically imbalanced, and there are few defective examples, we treat defect prediction as anomaly detection. We present our Reconstruction Error Probability Distribution (REPD) model which can handle point and collective anomalies. We compare it on five different traditional code feature datasets against five models: Gaussian Naive Bayes, logistic regression, k-nearest-neighbors, decision tree, and Hybrid SMOTE-Ensemble. In addition, REPD is compared on 24 semantic features datasets against previously mentioned models. In order to compare the performance of competing models, we utilize F1-score measure. By using statistical means, we show that our model produces significantly better results, improving F1-score up to 7.12%. Additionally, REPD’s robustness to dataset imbalance is analyzed by creating defect undersampled and non-defect oversampled datasets.