Mapping to optimal regions; A new method for multi-class classification task to reduce complexity

Article type: Research Article

Authors: Bavafaye Haghighi, Elham^; | Rahmati, Mohammad^; | Shiry Ghidary, Saeed | Palm, Günther

Affiliations: Computer Engineering & Information Technology Department, Amirkabir University of Technology, Tehran, Iran | Institute of Neural Information Processing, Ulm University, Ulm, Germany

Note: [] Corresponding author. E-mail: rahmati@aut.ac.ir

Abstract: Classification of data is an important problem which has attracted many researchers to introduce new approaches. In this paper, we propose Mapping to Optimal Regions (MOR) as a new method for multi-class classification task to reduce computational and memory complexities. It requires only one simple mapping from input space to optimal regions. The optimal domain is estimated using a multi objective cost function to increase the region size and the generalization ability of the mapping and to reduce the mapping error. Finally, the centers of optimal regions are determined with respect to the optimal size of the regions and the code assignment process which reduces the effect of inappropriate labeling. A Hierarchical version of MOR (HMOR) is presented for datasets with high number of classes or low dimensional feature spaces. By taking the advantage of MOR, the complexity reduces significantly in comparison to the other classifiers.

Keywords: Mapping to Optimal Regions (MOR), Hierarchical MOR (HMOR), code assignment, inappropriate labeling, reducing complexity, multi-class classification

DOI: 10.3233/AIC-140607

Journal: AI Communications, vol. 27, no. 4, pp. 387-404, 2014