Relationship between depth of decision trees and boosting performance

We have investigated strategies for enhancing ensemble learning algorithms for the analysis of high-dimensional biological data. Specifically we investigated strategies to force classifiers to consider the possible interactions between features. As a result an algorithm that induces decision trees with a feature non-replacement mechanism has been devised and tested on DNA microarray and proteomic datasets. The results show that feature non-replacement enables decision trees deeper than simple stumps to be used, thereby allowing feature interaction to be taken into account.