Table 7.

Statistical comparison of EBD and FI discretization methods

Evaluation Measure	Method	Mean (SEM)	Difference of Means	Z statistic (p-value)
C4.5 Accuracy	EBD	73.49% (2.07)	2.01	2.219
[0%, 100%]	FI	71.48% (2.12)		(0.026)
C4.5 AUC	EBD	73.22% (1.89)	1.07	2.732
[50%, 100%]	FI	72.15% (1.77)		(0.007)
C4.5 Robustness	EBD	72.55% (2.81)	-0.26	-0.261
[0%, ∞]	FI	72.81% (2.76)		(0.794)
NB Accuracy	EBD	77.55% (2.65)	0.76	2.080
[0%, 100%]	FI	76.79% (2.32)		(0.038)
NB AUC	EBD	74.83% (1.43)	1.11	2.711
[0%, 100%]	FI	73.71% (1.24)		(0.007)
NB Robustness	EBD	81.72% (2.92)	-0.68	-0.016
[50%, ∞]	FI	82.40% (2.59)		(0.987)
Stability	EBD	0.74 (0.025)	0.02	1.972
[0, 1]	FI	0.72 (0.029)		(0.049)
Mean # of intervals per predictor	EBD	1.27 (0.074)	0.11	1.686
[1, n]	FI	1.16 (0.038)		(0.092)

In the first column the range of a measure is given in square brackets where n is the number of instances in the dataset. In the last column the number on top in the last column is the Z statistic and the number at the bottom is the corresponding p-value. On all performance measures, except for the mean number of intervals per predictor, the Z statistic is positive when EBD performs better than FI. The two-tailed p-values of 0.05 or smaller are in bold, indicating that EBD performed statistically significantly better at that level.