An Approach to Training Decision Trees with the Relearning of Nodes
S. A. Mitrofanov, Eugene Semenkin
Abstract
Decision tree learning algorithms are quite old but very popular because of their efficiency. Decision trees are built sequentially from node to node, resulting in a "near-optimal" tree. This article discusses an approach to training a decision tree based on relearning already built nodes, i.e. a tree is not built sequentially, but with a return to the previous nodes. Such algorithms as Separation Measure and Differential Evolution are applied for this approach. Separation Measure selects the attributes to optimize the thresholds, and Differential Evolution performs optimization. The considered approach to learning a decision tree is compared with standard algorithms. Statistical analysis is performed to show the reliability of the results, with Student’s t-test being applied in this case. On average, the efficiency of the standard and modified decision tree learning algorithms are at the same level. However, a modified decision tree learning algorithm can find a tree that is more efficient than the trees built by the standard algorithm.