Fig. 3

More accurate decision boundaries are recovered using LANDMark models. Decision boundaries discovered by various classifiers on two-spirals dataset. The input data (a) was used to train (b) a single Extremely Randomized Tree, c a single decision tree, d, e two different LANDMark (Oracle) trees that demonstrate the randomness of the algorithm, f a single LANDMark (No Oracle) tree, g an Extremely Randomized Trees classifier consisting of 100 trees, h a Random Forest classifier consisting of 100 trees, i a LANDMark (Oracle) classifier consisting of 64 trees, and j a LANDMark (No Oracle) classifier consisting of 64 trees. Solid circles indicate data points used for training while crosses represent validation data. The balanced accuracy of each classifier is reported as the score. The shading in each plot is a qualitative representation of how confidently each model would predict the class of a particular sample. In panels b-f the red and blue regions are not shaded and represent where each model will predict either the red or blue spiral while in panels g–j the predictions of each ensemble member are averaged. In these panels darker regions represent areas where the prediction from each model is more confident