Tackling class overlap and imbalance problems in software defect prediction

Software defect prediction (SDP) is a promising solution to save time and cost in the software testing phase for improving software quality. Numerous machine learning approaches have proven effective in SDP. However, the unbalanced class distribution in SDP datasets could be a problem for some conve...

Full description

Saved in:
Bibliographic Details
Published in:Software quality journal 2018-03, Vol.26 (1), p.97-125
Main Authors: Chen, Lin, Fang, Bin, Shang, Zhaowei, Tang, Yuanyan
Format: Article
Language:English
Subjects:
Classifiers
Cleaning
Compilers
Computer Science
Data Structures and Information Theory
Datasets
Defects
Interpreters
Machine learning
Operating Systems
Programming Languages
Software
Software Engineering/Programming and Operating Systems
Software testing
Statistical analysis
Statistical methods
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Software defect prediction (SDP) is a promising solution to save time and cost in the software testing phase for improving software quality. Numerous machine learning approaches have proven effective in SDP. However, the unbalanced class distribution in SDP datasets could be a problem for some conventional learning methods. In addition, class overlap increases the difficulty for the predictors to learn the defective class accurately. In this study, we propose a new SDP model which combines class overlap reduction and ensemble imbalance learning to improve defect prediction. First, the neighbor cleaning method is applied to remove the overlapping non-defective samples. The whole dataset is then randomly under-sampled several times to generate balanced subsets so that multiple classifiers can be trained on these data. Finally, these individual classifiers are assembled with the AdaBoost mechanism to build the final prediction model. In the experiments, we investigated nine highly unbalanced datasets selected from a public software repository and confirmed that the high rate of overlap between classes existed in SDP data. We assessed the performance of our proposed model by comparing it with other state-of-the-art methods including conventional SDP models, imbalance learning and data cleaning methods. Test results and statistical analysis show that the proposed model provides more reasonable defect prediction results and performs best in terms of G-mean and AUC among all tested models.
ISSN:0963-9314
1573-1367
DOI:10.1007/s11219-016-9342-6