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

Table 10

Average predictive accuracy of extracted rulesets from DIMLP-Bs, DIMLP-As, and single DIMLPs (columns 2, 3, and 4). Average complexity of rulesets (last three columns).

Dataset	DIMLP-B	DIMLP-A	DIMLP	DIMLP-B	DIMLP-A	DIMLP
Australian Credit Appr.	86.5	84.9	86.1	22.7 3.7	82.7 5.1	30.0 3.9
Breast Cancer	96.5	96.2	96.4	12.5 2.7	25.2 3.6	13.3 2.8
Breast Cancer 2	95.6	95.8	95.9	20.8 3.1	27.2 3.3	19.9 3.1
Breast Canc. (prognostic)	79.0	77.7	77.7	12.6 2.8	24.6 3.5	15.0 2.8
Bupa Liver Disorders	70.9	67.2	67.8	36.9 3.3	31.8 3.2	37.7 3.5
Chess (kr-versus-kp)	99.5	99.7	99.4	32.5 4.0	36.3 4.2	35.7 4.2
Coronary Heart Disease	91.6	92.3	90.9	44.8 4.0	71.6 4.6	53.1 4.2
German Credit	73.6	72.6	73.9	56.7 4.3	93.9 5.1	73.8 5.0
Glass (binary)	77.8	81.1	79.5	13.7 2.7	19.9 3.2	14.5 2.8
Haberman	74.3	73.3	72.8	7.8 1.9	2.3 0.6	6.7 1.8
Heart Disease	84.3	80.5	81.8	20.6 3.2	39.3 4.2	23.5 3.4
ILPD (liver)	70.7	70.8	68.8	23.9 3.1	25.2 2.8	31.1 3.2
Ionosphere	92.1	90.6	92.1	19.3 2.9	29.5 3.2	20.5 3.1
Istanbul Stock Exch.	77.2	75.1	76.0	21.4 2.8	23.4 2.9	26.0 3.0
Labor	84.3	87.4	86.1	7.3 2.2	9.2 2.6	7.2 2.2
Musk1	85.8	86.0	86.2	57.4 4.4	90.1 4.3	60.1 4.3
Pima Indians	76.3	74.2	75.5	38.8 3.3	47.0 3.6	42.3 3.6
Promoters	83.0	81.1	81.3	11.8 2.7	20.4 3.4	12.4 2.8
Saheart	71.9	68.6	70.9	29.2 3.3	27.5 3.3	29.2 3.4
Sonar	79.0	78.4	77.9	24.1 3.2	40.2 3.8	24.5 3.2
Spect heart	72.2	67.9	69.2	20.4 3.2	26.8 3.6	24.4 3.3
Splice junct.	95.1	95.3	94.6	108.3 5.5	172.6 6.4	124.6 7.7
Svmguide	96.8	96.9	96.8	38.0 2.9	69.8 3.2	38.7 2.9
Tictactoe	98.4	98.7	98.5	27.3 3.7	32.8 3.8	31.7 3.9
Vertebral Column	84.0	82.7	83.4	16.7 2.7	27.8 3.2	17.3 2.8