DBDNN-Estimator: A Cross-Project Number of Fault Estimation Technique

Abstract

Cross-project fault prediction (CPFP) uses data sets from projects to predict faulty/non-faulty modules. Cross-project fault number estimation (CPFNE) is one step ahead of CPFP, because it not only predicts faulty modules but also estimates the number of faults in that module. In this article, we proposed a new computational architecture using a deep belief network and deep neural network called DBDNN-Estimator for CPFNE. We investigated the effectiveness of our proposed approach on five projects and their respective versions from the PROMISE repository in our experiment and compared its performance over the existing eight benchmark approaches. We found that the proposed model required a few instances from the source project for optimal performance. Out of 23, we found that DBDNN-Estimator significantly outperforms in 19 and 14 data sets over baseline approaches in terms of mean absolute error (MAE) and mean squared error (MSE), respectively. The mean MAE and MSE produced by the proposed work are 0.38 ± 0.023 and 2.29 ± 0.18 , respectively, which is minimum amongst benchmark techniques. We also found the Kendall and Fault Percentage Average (FPA) of the proposed model significantly better than baseline methods in 17 projects. We found the DBDNN-Estimator produces optimal results for small, moderate, and large-size software projects. The model is stable and tackles class imbalance and overfitting problems. Graphical Abstract: [Figure not available: see fulltext.] © 2023, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.