Mixed decision trees: An evolutionary approach

In the paper, a new evolutionary algorithm (EA) for mixed tree learning is proposed. In non-terminal nodes of a mixed decision tree different types of tests can be placed, ranging from a typical univariate inequality test up to a multivariate test based on a splitting hyperplane. In contrast to classical top-down methods, our system searches for an optimal tree in a global manner, i.e. it learns a tree structure and tests in one run of the EA. Specialized genetic operators allow for generating new sub-trees, pruning existing ones as well as changing the node type and the tests. The proposed approach was experimentally verified on both artificial and real-life data and preliminary results are promising.