Notable changes:
1) Supports classification and regression
2) 3 classification criteria, 1 regression criterion
3) A new dataset is provided to test regression (Boston House Prices)
4) Weights are removed from the algorithm entirely. If the need for weights can be justified, I would welcome reintroducing them, but for the refactoring I left them out.
5) The subset of dimensions (F) to split on is fixed for the entire tree, not at each node. This is more in line with CART and RandomForests.
6) A max_depth parameter is offered to limit the size of the constructed trees.
7) Randomisation is fixed with python's random module, but can be seeded.
8) For classification, the number of classes must be provided when the tree is constructed. This is because the tree cannot necessarily infer the correct number of labels at the time of training if only a subset of the data is used for individual trees.
9) For classification, labels are not normalised internally. Labels must be provided to the algorithm in the range [0, ..., K)
10) For classification, the leaf nodes retain the distribution of classes. This means that it is possible to query the tree for the probability distribution of a test sample
Brian Holt