Evaluating the Impact of Data Transformation Techniques on the Performance and Interpretability of Software Defect Prediction Models

Table 3

The p-values of the Wilcoxon signed-rank test and average percentage performance gains on all six performance evaluation indicators between the transformed models and original models built with different classifiers.

p-Values and percentage	O vs. L		O vs. M		O vs. Z
RF-Accuracy		0		0		0
RF-Precision		−0.3%		0		0
RF-Recall		0.7%		0		0
RF-F1		0.2%		0		0
RF-AUC		0		0		0
RF-MCC		0		0		0
LR-Accuracy		5%		3%		3%
LR-Precision		4%		6%		4%
LR-Recall		7%		−2%		2%
LR-F1		7%		0		4%
LR-AUC		5%		2%		3%
LR-MCC		2%		12%		15%
DT-Accuracy		1%		0.9%		0.9%
DT-Precision		0.9%		0.9%		0.9%
DT-Recall		5%		4%		4%
DT-F1		3%		3%		3%
DT-AUC		2%		2%		2%
DT-MCC		8%		8%		7%
NB-Accuracy		7%		2%		2%
NB-Precision		−5%		3%		3%
NB-Recall		61%		4%		4%
NB-F1		31%		5%		5%
NB-AUC		7%		2%		2%
NB-MCC		26%		9%		9%
KNN-Accuracy		3%		1%		2%
KNN-Precision		3%		4%		4%
KNN-Recall		7%		−0.8%		0
KNN-F1		5%		1%		2%
KNN-AUC		4%		2%		2%
KNN-MCC		17%		9%		12%
MLP-Accuracy		2%		4%		4%
MLP-Precision		1%		4%		4%
MLP-Recall		7%		3%		7%
MLP-F1		5%		4%		7%
MLP-AUC		2%		3%		4%
MLP-MCC		8%		13%		17%