Handling Imbalanced Data Sets in Multistage Classification

Abstract

Multistage classification is a logical approach, based on a divide-and-conquer solution, for dealing with problems with a high number of classes. The classification problem is divided into several sequential steps, each one associated to a single classifier that works with subgroups of the original classes. In each level, the current set of classes is split into smaller subgroups of classes until they (the subgroups) are composed of only one class. The resulting chain of classifiers can be represented as a tree, which (1) simplifies the classification process by using fewer categories in each classifier and (2) makes it possible to combine several algorithms or use different attributes in each stage. Most of the classification algorithms can be biased in the sense of selecting the most populated class in overlapping areas of the input space. This can degrade a multistage classifier performance if the training set sample frequencies do not reflect the real prevalence in the population. Several techniques such as applying prior probabilities, assigning weights to the classes, or replicating instances have been developed to overcome this handicap. Most of them are designed for two-class (accept-reject) problems. In this article, we evaluate several of these techniques as applied to multistage classification and analyze how they can be useful for astronomy. We compare the results obtained by classifying a data set based on Hipparcos with and without these methods.