Mdl = treetrain (X, Y, W, opts) grows a tree on the NxP predictor matrix X and returns it as a structure with one row per node. It is the fitting engine shared by ClassificationTree and RegressionTree, and is not meant to be called directly.

Y is an Nx1 vector of class indices in 1:K when fitting a classifier and the response itself when fitting a regression. W is an Nx1 vector of observation weights, already adjusted for any prior.

opts is a structure carrying NumClasses, MinParent, MinLeaf, MaxSplits, SplitCriterion, MergeLeaves and, for a regression, QEToler. SplitCriterion is "gdi" or "deviance" for a classifier and "mse" for a regression, and it is what selects between the two.

NumVariablesToSample, when present, is the number of predictors each node chooses its split from, drawn afresh at every node, and Seed, an integer from 0 to 2^32-1, seeds that draw, so the same seed grows the same tree. A node whose predictors hold no valid split is a leaf. Without NumVariablesToSample every predictor is tried and nothing is drawn.

CategoricalPredictors, when present, lists the columns of X that hold levels rather than numbers, and such a column is split into two sets of levels. A regression orders the levels by mean response and two classes by the probability of the first class. More classes search every partition when the node holds at most MaxNumCategories levels, 10 by default, and otherwise take MATLAB’s heuristics, as AlgorithmForCategorical selects: "auto" (default), "exact", "pullleft", "pca" or "ovabyclass".

The returned structure holds Children, Parent, CutPredictorIndex, CutPoint, CutCategories, IsBranchNode, NodeSize, NodeWeight and NumNodes, plus ClassWeight and ClassCount for a classifier or NodeMean and NodeError for a regression. Nodes are numbered as they are created, so a parent always precedes its children.

The cost complexity pruning sequence is not built here. A tree outlives the data it was grown on, so the sequence is __treeprune__, which takes a node table and the risk the caller measures by.

See also: treepredict

Source Code: treetrain