A Hybrid Approach Using Oversampling Technique and Cost-Sensitive Learning for Bankruptcy Prediction

Table 3

The overall results of all experimental approaches for KRBDS.

No	Method	Resample approach	Classifier	AUC	G-mean	Average Rank	p-value
1	BG	None	Bagging	78.8±0.4	70.8±0.8	9.0	3.9×10^−5
2	AB	None	AdaBoost	84.9±0.8	78.2±0.6	7.0	0.0023
3	RF	None	Random Forest	86.2±0.6	79.9±0.6	4.7	0.069
4	MLP	None	MLP	86.7±0.8	80.1±1.0	2.6	0.487
5	USC-BG	Under-sampling method based on clustering technique (USC) [43]	Bagging	65.1±1.6	53.6±4.9	11.2	1.2×10^−7
6	USC-AB		AdaBoost	59.7±3.0	56.3±5.0	12.9	5.6×10^−10
7	USC-RF		Random Forest	64.7±1.0	62.6±1.9	11.9	1.5×10^−8
8	USC-MLP		MLP	46.9±2.7	36.5±3.7	14.0	1.1×10^−11
9	OSE-BG	Oversampling method using SMOTE-ENN (OSE) [41]	Bagging	83.9±0.3	77.4±0.3	7.8	5.1×10^−4
10	OSE-AB		AdaBoost	85.4±0.7	78.5±0.4	6.2	0.009
11	OSE-RF		Random Forest	86.6±0.7	80.2±1.0	3.3	0.285
12	OSE-MLP		MLP	72.8±2.1	69.8±1.8	10.0	3.3×10^−6
13	RFCI [42]	Under-sampling method using IHT concept	CBoost	86.6±0.7	79.1±3.5	3.1	0.336
14	HAOC	Oversampling method using SMOTE-ENN (with balancing ratio = 0.08)	CBoost	87.1±0.6	81.1±0.8	1.3	-