Fig. 4

The tree-like Bayesian Structure Learning Algorithm can select a root for structure learning in tree-like or non-tree-like networks. Roots were selected automatically for two representative networks across a range of sample size limitations: a the tree-like Child network and b the non-tree-like Alarm network. Any node ever selected as a root has a red outline, where increasing line width indicates increasing frequency of selection as a root. Nodes never selected as a root have blue outlines of fixed width. c A quantitative summary of the root nodes selected, as a function of sample size. Selection from a tree-like structure is straightforward and consistent; from a non-tree-like structure there is increased variability, but reasonable roots (excluding directionality) are typically chosen. Feature 24 was used as the root for previous Alarm network learning work. It is worth noting that selection of a better root could improve the TL-BSLA’s TPR and PPV even further