作者: Tin Kam Ho
关键词:
摘要: Using a number of measures for characterising the complexity classification problems, we studied comparative advantages two methods constructing decision forests – bootstrapping and random subspaces. We investigated collection 392 two-class problems from UCI depository, observed that there are strong correlations between classifier accuracies length class boundaries, thickness manifolds, nonlinearities boundaries. found characteristics both difficult easy cases where combination no better than single classifiers. Also, method is when training samples sparse, subspace classes compact boundaries smooth.