Class noise detection by multiple voting

Ensemble learning has been used for identifying and eliminating mislabeled instances. Its main idea is to use a set of learning algorithms to create classifiers which serve as noise filters. The diversities of classifiers lead to the diversities of their judgments on noises. Voting mechanism is used to fuse these different judgments and make final decisions on which data are noises. By making use of diversities among classifiers, this voting based method has shown better performance than single classifier. Although many different types of voting based noise detection methods have been proposed (e.g., majority voting, consensus voting), these methods conduct voting only for one time. This one time voting mechanism is biased to the distribution of data that are selected for training ensemble classifiers. To reduce this bias, we propose to use multiple voting for noise detection. The design of multiple voting is straightforward. Through both theoretical and experimental analysis, we find that multiple voting can detect noises more accurately than single voting.