Construct Rough Decision Forests Based on Sequentially Data Reduction

Decision forests have been proven to be a promising technique to improve classification performance. The improvement comes from the diversity among the individual classifiers. It is believed that diverse ensembles have a good potential for improving the accuracy compared with non-diverse ensembles. In this paper, we propose a technique to construct diverse decision forests based on rough set reduction. The method recursively generates a sequence of minimal reducts as the training subspaces, where each minimal reduct is extracted from the rest attribute set, the attributes contained in the former minimal reducts will be removed in extracting the new reduct. Therefore, there is no common attribute in all the reducts created by this technique, which guarantees the decision trees trained by distinct reducts reflects different classification information of the training set. Final decisions are made based on outputs from the decision trees by the majority-voting rule. Experiments show the proposed method gets a good performance