Ordinal Classification of Imbalanced Data with Application in Emergency and Disaster Information Services

Abstract

This paper considers the problem of ordinal classification of imbalanced data, i.e., the class distribution is imbalanced among the classes, and the classes have ordered class labels. There has been an increasing focus on the ordinal classification problem, and various methods for ordinal classification have been proposed. However, the previous studies implicitly assumed the balanced data sets; if the dataset is imbalanced, they tend to be biased towards the majority class. This paper proposes a new method for ordinal classification of imbalanced data, called the alpha-weighted k-nearest neighbors method (alpha-wkNN). In the framework of the weighted k-nearest neighbors method, alpha-wkNN determines the class membership using the alpha quantile of the estimated class probability distribution, where the value of alpha is determined such that the impact of the class imbalance is mitigated. The effectiveness of the proposed method is demonstrated using a real dataset of text documents from an emergency and disaster information service provider: an event's risk level on a discrete rating scale is predicted using the event descriptions written in natural language.