Active Learning for Ordinal Classification Based on Adaptive Diversity-Based Uncertainty Sampling

Abstract

Although a plethora of research has been conducted on active learning, little research attention has been focused on active learning for ordinal classification. Traditional multi-class active learning methods are typically designed for nominal multi-class classification. Therefore, they usually perform unsatisfactorily in ordinal classification settings. In ordinal classification, the cost of misclassifying an instance into an adjacent class is naturally lower than that of misclassifying it into a more disparate class. This principle is called ordering information. However, traditional active learning methods typically do not consider this ordering information during query selection. This paper proposes a novel adaptive hybrid active learning method for ordinal classification by considering the ordering information. In the proposed method, an uncertainty measure is introduced to select the hard-to-predict instances distributed between adjacent classes. In addition, a diversity measure is incorporated with the uncertainty measure to alleviate the potential sampling redundancy. Finally, an expected cost minimization measure with ordering information is designed. This measure balances the contributions of the uncertainty and diversity measures and prompts the algorithm to select the instances most likely to decrease the misclassification cost of the model. Extensive experiments on eleven public ordinal datasets demonstrate the superiority of the proposed method over several state-of-the-art methods.