Empirical study on virtual order of class labels in nominal classification

Chengwei Li, Han Liu, Zhong Ming

Abstract: Binary Decomposition can be adopted in ordered and unordered ways. Inspired by the case that label order information can be exploited to improve the ordinal classification performance through adopting an ordered decomposition strategy, this paper explores whether the effectiveness of binary decomposition in nominal classification tasks can be improved by setting a virtual label order. The essential purpose of setting a virtual order is to obtain small intra-class distances and large inter-class distances after binary decomposition, such that simpler binary classification tasks are obtained. The experimental results show that setting a virtual order results in an improvement of the classification performance as expected. However, the performance obtained by setting a virtual order does not show significant superiority in comparison with the one produced by some unordered decomposition strategies, and the reasons have been analysed in the context of the relationship between virtual orders and inter-class distances.