Switching class labels to generate classification ensembles

摘要: Ensembles that combine the decisions of classifiers generated by using perturbed versions training set where classes examples are randomly switched can produce a significant error reduction, provided large numbers units and high class switching rates used. The this procedure have statistically uncorrelated errors in set. Hence, ensembles they form exhibit similar dependence on ensemble size, independently classification problem. In particular, for binary problems, performance data be analysed terms Bernoulli process. Experiments several UCI datasets demonstrate improvements accuracy obtained these class-switching ensembles.