An Empirical Study for Enhanced Software Defect Prediction Using a Learning-Based Framework

The object of software defect prediction (SDP) is to identify defect-prone modules. This is achieved through constructing prediction models using datasets obtained by mining software historical depositories. However, data mined from these depositories are often associated with high dimensionality, class imbalance, and mislabels which deteriorate classification performance and increase model complexity. In order to mitigate the consequences, this paper proposes an integrated preprocessing framework in which feature selection (FS), data balance (DB), and noise filtering (NF) techniques are fused to deal with the factors that deteriorate learning performance. We apply the proposed framework on three software metrics, namely static code metric (SCM), object oriented metric (OOM), and combined metric (CombM) and build models based on four scenarios (S): (S1) original data; (S2) FS subsets; (S3) FS subsets after DB using random under sampling (RUS) and synthetic minority oversampling technique (SMOTE); (S4) FS subsets after DB (RUS and SMOTE); and NF using iterative partitioning filter (IPF) and iterative noise filtering based on the fusing of classifiers (INFFC). Empirical results show that 1. the integrated preprocessing of FS, DB, and NF improves the performance of all the models built for SDP, 2. for all FS methods, all the models improve performance progressively from S2 through to S4 in all the software metrics, 3. model performance based on S4 is statistically significantly better than the performance based on S3 for all the software metrics, and 4. in order to achieve optimal model performance for SDP, appropriate implementation of the proposed framework is required. The results also validate the effectiveness of our proposal and provide guidelines for achieving quality training data that enhances model performance for SDP.