Table 4.

A comparison between cross-feature analysis (CFA) and a version of FRaC that models numeric features as discrete value ranges to measure the extent to which the use of surprisal affects anomaly detection

Feature predictor	CFA	Discrete FRaC	p-value
Naïve Bayes	13	29	0.00258
RIPPER	10	30	0.000416
Decision tree	12	28	0.00203
Linear Kernel SVM	15	25	0.13
RBF Kernel SVM	14	19	0.376
Tree, linear and RBF SVM combined	9	25	0.00178

The second and third columns indicate the number of data sets from Table 1 with superior AUC for each feature model type, and the fourth column shows the p-value from a one-tailed paired t-test comparing the AUC scores across all data sets. (Due to ties, neither of the two compared methods has the superior AUC for some data sets.)