Heuristic Settings of Virtual Orders of Class Labels for Nominal Classification

Han Liu, Yinghui Pan, Hao Chen

Abstract: In multi-class classification tasks, binary decomposition of the class attribute is necessary for some learning algorithms such as support vector machines. While the class attribute is nominal, the binary decomposition is typically done in an unordered way, which can result in a negative impact on the effectiveness and efficiency of learning binary classifiers in some scenarios. In this paper, we explore whether it is achievable to improve the performance of nominal classification by setting virtual class orders for achieving ordered binary decomposition of the class attribute. Specifically, we propose a framework to search for virtual label orders that are suitable for effectively transforming a nominal classification problem into simpler binary ones, where the framework consists of a heuristic search module for searching local optimal orders and an ensemble module for combining the orders. The experimental results show that our framework leads to an improvement of the classification performance in comparison with arbitrarily assigning a label order or setting the class labels to be unordered. The current work can be regarded as a type of class representation learning.