Error rate


cost = treetest(t,'resubstitution')
cost = treetest(t,'test',X,y)
cost = treetest(t,'crossvalidate',X,y)
[cost,secost,ntnodes,bestlevel] = treetest(...)
[...] = treetest(...,param1,val1,param2,val2,...)

    Note:   treetest will be removed in a future release. Use classregtree.test instead.


    Note:   This function is superseded by the test method of the classregtree class and is maintained only for backwards compatibility. It accepts objects t created with the classregtree constructor.

cost = treetest(t,'resubstitution') computes the cost of the tree t using a resubstitution method. t is a decision tree as created by the treefit function. The cost of the tree is the sum over all terminal nodes of the estimated probability of that node times the node's cost. If t is a classification tree, the cost of a node is the sum of the misclassification costs of the observations in that node. If t is a regression tree, the cost of a node is the average squared error over the observations in that node. cost is a vector of cost values for each subtree in the optimal pruning sequence for t. The resubstitution cost is based on the same sample that was used to create the original tree, so it underestimates the likely cost of applying the tree to new data.

cost = treetest(t,'test',X,y) uses the predictor matrix X and response y as a test sample, applies the decision tree t to that sample, and returns a vector cost of cost values computed for the test sample. X and y should not be the same as the learning sample, which is the sample that was used to fit the tree t.

cost = treetest(t,'crossvalidate',X,y) uses 10-fold cross-validation to compute the cost vector. X and y should be the learning sample, which is the sample that was used to fit the tree t. The function partitions the sample into 10 subsamples, chosen randomly but with roughly equal size. For classification trees, the subsamples also have roughly the same class proportions. For each subsample, treetest fits a tree to the remaining data and uses it to predict the subsample. It pools the information from all subsamples to compute the cost for the whole sample.

[cost,secost,ntnodes,bestlevel] = treetest(...) also returns the vector secost containing the standard error of each cost value, the vector ntnodes containing number of terminal nodes for each subtree, and the scalar bestlevel containing the estimated best level of pruning. bestlevel = 0 means no pruning, i.e., the full unpruned tree. The best level is the one that produces the smallest tree that is within one standard error of the minimum-cost subtree.

[...] = treetest(...,param1,val1,param2,val2,...) specifies optional parameter name-value pairs chosen from the following table.


The number of cross-validations samples (default is 10).


Either 'se' (default) to choose the smallest tree whose cost is within one standard error of the minimum cost, or 'min' to choose the minimal cost tree.


Find the best tree for Fisher's iris data using cross-validation. The solid line shows the estimated cost for each tree size, the dashed line marks one standard error above the minimum, and the square marks the smallest tree under the dashed line.

% Start with a large tree.
load fisheriris;
t = treefit(meas,species','splitmin',5);

% Find the minimum-cost tree.
[c,s,n,best] = treetest(t,'cross',meas,species);
tmin = treeprune(t,'level',best);

% Plot smallest tree within 1 std of minimum cost tree.
[mincost,minloc] = min(c);
xlabel('Tree size (number of terminal nodes)')


[1] Breiman, L., J. Friedman, R. Olshen, and C. Stone. Classification and Regression Trees. Boca Raton, FL: CRC Press, 1984.

See Also


Was this topic helpful?