Ant-inspired algorithms for decision tree induction: An evaluation on biomedical signals

Abstract

In this paper we present an evaluation of ant-inspired method called ACO_DTree over biomedical data. The algorithm maintains and evolves a population of decision trees induced from data. The core of the algorithm is inspired by the Min-Max Ant System. In order to increase the speed of the algorithm we have introduced a local optimization phase. The generalization ability has been improved using error based pruning of the solutions. After parameter tuning, we have conducted experimental evaluation of the ACO_DTree method over the total of 32 different datasets versus 41 distinct classifiers. We conducted 10-fold crossvalidation and for each experiment obtained about 20 quantitative objective measures. The averaged and best-so-far values of the selected measures (precision, recall, f-measure,...) have been statistically evaluated using Friedman test with Holm and Hochberg post-hoc procedures (on the levels of α = 0.05 and α = 0.10). The ACO_DTree algorithm performed significantly better (α = 0.05) in 29 test cases for the averaged f-measure and in 14 cases for the best-so-far f-measure. The best results have been obtained for various subsets of the UCI database and for the dataset combining cardiotocography data and data of myocardial infarction.