Affiliations: [a] School of Information Science and Technology, Southwest Jiaotong University, Chengdu, China
| [b] School of Transport and Logistics Engineering, Southwest Jiaotong University, Chengdu, China
| [c] Department of Information Technology, College of Computer Science and Information Technology, Karary University, Omdurman, Sudan
Correspondence:
[*]
Corresponding author. Tianrui Li, School of Information Science and Technology, Southwest Jiaotong University, Chengdu, China. E-mail: [email protected].
Abstract: The object of Software Defect Prediction (SDP) is to identify modules that are prone to defect. This is achieved by training prediction models with datasets obtained by mining software historical depositories. When one acquires data through this approach, it often includes class imbalance which has an unequal class representation among their example. We hypothesize that the imbalance learning is not a problem in itself and decrease in performance is also influenced by other factors related to class distribution in the data. One of these is the existence of noisy and borderline examples. Thus, the objective of our research is to propose a novel preprocessing method using Synthetic Minority Over-Sampling Technique (SMOTE), Fuzzy-rough Instance Selection type II (FRIS-II) and Iterative Noise Filter based on the Fusion of Classifiers (INFFC) which can overcome these problems. The experimental results show that the new proposal significantly outperformed all the methods compared in this study.
