Evolutionary design of decision trees

Throughout fifteen years of research we designed, developed and extensively evaluated the process of building decision trees with evolutionary algorithms, which was applied to a variety of real-world problems in the fields of medicine, software engineering and telecommunications. Decision trees are one of the most widespread and commonly used machine learning method, which enables straightforward and transparent classification, with the possibility of validation of constructed knowledge models. We replaced the traditional greedy top-down induction approach with an original evolutionary algorithm. In this manner, we succeeded to improve the predictive performance, while reducing the complexity of constructed models and achieving significantly higher classification balance. With the introduction of expert knowledge into an iterative process of building the solution we designed an innovative algorithm that covers larger search space of potential solutions than traditional methods, which led to the discovery of a potentially new knowledge in the domain of cardiovascular problems within the younger population. Further original scientific contributions include self-adaptive construction process, multi-level hierarchical decomposition of a learning set, vector decision trees, autonomous evolutionary process, smart crossover operator, similarity based selection and multi-population evolutionary algorithm.

Published in:

Vili Podgorelec, Matej Šprogar, Sandi Pohorec, Evolutionary design of decision trees, WIREs Data Mining and Knowledge Discovery, Elsevier, 3(2), pp. 63-82, 2013.

Comments are closed.