Uncertainty-driven ensemble classification exploiting unlabeled data

Abstract

This works investigates the use of margin and diversity, two key concepts in ensemble learning, to develop a versatile uncertainty-driven ensemble classifier, under the scarcity of labeled data. New semi-supervised definitions are proposed for both margin and diversity. To strengthen the semi-supervised classification scheme, classification accuracy estimates are accompanied by new robust ensemble metrics. These measures reflect the degree of certainty associated with the classification outcome of each pattern. They are calculated in a semi-supervised way. The proposed metrics are exploited in an original uncertainty-driven fusion process that combines multiple classification results. The relevance of these new criteria is examined in change detection experiments. The underlying fusion rule significantly improves the change detection performance.