Constructing Rough Decision Forests

Decision forests are a type of classification paradigm which combines a collection of decision trees for a classification task, instead of depending on a single tree. Improvement of accuracy and stability is observed in experiments and applications. Some novel techniques to construct decision forests are proposed based on rough set reduction in this paper. As there are a lot of reducts for some data sets, a series of decision trees can be trained with different reducts. Three methods to select decision trees or reducts are presented, and decisions from selected trees are fused with the plurality voting rule. The experiments show that random selection is the worst solution in the proposed methods. It is also found that input diversity maximization doesn't guarantee output diversity maximization. Hence it cannot guarantee a good classification performance in practice. Genetic algorithm based selective rough decision forests consistently get good classification accuracies compared with a single tree trained by raw data as well as the other two forest constructing methods.