Performance of various classifiers on a simple example dataset, discretization/data_gen2_train.csv and discretization/data_gen2_test.csv, trained and tested on separate data generated from a BN of 5 discrete and 15 continuous variables (n = 25), and also averages over 10 similar, randomly generated networks.

<p>This illustrates the differences between CGBayesNets and two other software packages, BNfinder 2.0 and Weka 3.6.9. Training data is presented in two forms: in its original form, and in its discretized form, where continuous nodes are binned into 10 equal-width discrete bins. The “nodes” columns refer to the number of nodes in the Markov blanket of the phenotype node in the discovered network. “AUC” is the Area Under receiver-operator characteristic Curve, evaluated by predicting the phenotype node given values of the other 19 nodes in a separate test set. “BNfinder 2.0 (K2)” refers to BNfinder 2.0 supplied with node-parent constraints consistent with the topological ordering of nodes in the true network. “BNfinder 2.0 (reverse-K2)” indicates BNfinder 2.0 supplied with node-parent constraints consistent with the opposite of the topological ordering of the nodes in the true network. No data is reported for Weka 3.6.9 on the Original data, since Weka 3.6.9 does not handle Bayesian Networks with continuous data variables. Average AUCs over 10 similarly-generated networks are shown for each method, before and after discretization.