A Comparison Study on Rule Extraction from Neural Network Ensembles, Boosted Shallow Trees, and SVMs

Table 2

Comparison between the average predictive accuracy obtained by the best model (column three) and the average predictive accuracy of the best extracted rulesets (column five). The last three columns indicate the average fidelity, the average number of generated rules, and the average number of antecedents per rule, respectively.

Dataset	Model	Avg. Acc.	Model	Avg. rules Acc.	Diff.	Fid.	#rules	#Ant.
Australian Credit Appr.	DIMLP-B	86.7	DIMLP-B	86.5	0.2	97.9	21.4	2.8
Breast Cancer	QSVM-G	97.2	QSVM-G	96.7	0.5	98.7	11.6	2.9
Breast Cancer 2	BST-G1	97.5	BST-G1	96.4	1.1	97.6	31.6	3.6
Breast Canc. (prognostic)	DIMLP-B	81.0	DIMLP-B	79.0	2.0	94.8	12.6	2.8
Bupa Liver Disorders	DIMLP-B	72.7	BST-M4	71.1	1.6	91.8	37.2	3.4
Chess (kr-versus-kp)	BST-G3	99.6	BST-G3	99.8		99.7	37.0	4.3
Coronary Heart Disease	DIMLP-A	94.6	DIMLP-A	92.3	2.3	96.2	71.6	4.6
German Credit	QSVM-P3	75.9	QSVM-P3	75.1	0.8	94.4	85.4	5.4
Glass (binary)	BST-G4	88.2	BST-M3	85.4	2.8	95.8	16.7	3.2
Haberman	BST-M1	75.0	BST-M1	75.0	0.0	100.0	3.0	1.3
Heart Disease	DIMLP-B	85.8	DIMLP-B	84.3	1.5	95.2	20.6	3.2
ILPD (liver)	BST-G3	70.8	DIMLP-A	70.8	0.0	96.8	25.2	2.8
Ionosphere	QSVM-G	94.4	BST-M1	92.1	2.3	97.8	14.3	2.8
Istanbul Stock Exch.	QSVM-L	77.7	QSVM-L	77.6	0.1	94.4	30.0	3.0
Labor	QSVM-G	95.1	BST-R4	89.6	5.5	95.4	9.8	2.5
Musk1	DIMLP-A	94.1	BST-G4	87.2	6.9	91.7	78.1	4.5
Pima Indians	QSVM-G	76.8	DIMLP-B	76.3	0.5	97.0	38.8	3.3
Promoters	BST-G3	92.1	BST-M1	84.7	7.4	90.7	11.4	2.7
Saheart	DIMLP-B	72.7	BST-M3	72.3	0.4	97.5	26.3	3.4
Sonar	BST-G4	88.6	BST-M3	81.9	6.7	88.3	35.8	4.2
Spect Heart	DIMLP-B	73.0	DIMLP-B	72.2	0.8	94.8	20.4	3.2
Splice junct.	BST-M4	97.3	BST-M4	97.1	0.2	99.1	63.1	4.7
Svmguide	BST-M4	97.3	BST-M4	97.2	0.1	99.7	44.5	3.1
Tictactoe	BST-G4	99.9	BST-G4	100		99.9	28.6	4.0
Vertebral Column	DIMLP-B	85.5	DIMLP-B	84.0	1.5	96.1	16.7	2.7