Evaluation of sampling techniques in software fault prediction using metrics and code smells

The highly imbalanced nature of software fault datasets results in poor performance of machine leaning techniques used for software fault prediction. The objective of this paper is to evaluate sampling techniques and Meta-Cost learning in software fault prediction to alleviate problem of imbalanced data. We evaluate four sampling techniques in metrics as well as code smells based fault prediction on fault data sets of two open source systems ANT and POI. Our results indicate that Resample technique is best for metrics based fault prediction whereas Synthetic Minority Oversampling is best suited for code smells based fault prediction. The results are presented in terms of accuracy measures like G-Mean, F-measure and area under ROC curve. We also evaluate Meta-Cost learning and found that all sampling techniques outperform Meta-Cost learning. Our results also indicate that software metrics are better predictor of software faults than code smells.