What are the limits of evolutionary induction of decision trees?

Abstract

For typical assessment of applying machine learning or data mining techniques, accuracy and interpretability are usually the most important elements. However, when the analyst is faced with real contemporary big data problems, scalability and efficiency become crucial factors. Parallel and distributed processing support is often an indispensable component of operational solutions. In the paper, we investigate the applicability of evolutionary induction of decision trees to large-scale data. We focus on the existing Global Decision Tree system, which searches the tree structure and tests in one run of an evolutionary algorithm. Evolved individuals are not encoded, so the specialized genetic operators and their application schemes are used. As in most evolutionary data mining systems, every fitness evaluation needs processing the whole training dataset. For high-dimensional datasets, this operation is very time consuming and to overcome this deficiency, two acceleration solutions, based on the most promising, latest approaches (NVIDIA CUDA and Apache Spark) are presented. The fitness calculations are delegated, while the core evolution is unchanged. In the experimental part, among others, we identify what are dataset dimensions which can be efficiently processed in the fixed time interval.