Two-stage SVMs for Solving Multi-class Problems

The conventional pairwise classification shows superior performances for those classifiable samples, but unclassifiable regions exist. DDAG based SVMs can resolve unclassified regions, but it only employ one decision function at each step, and never considers all decision function together, which may hurt its classification performance. In this paper we propose two-stage SVMs to combine their merits together, which resolve unclassifiable regions and keep the classification results same as conventional pairwise classification for the data in classifiable regions. To classify the data in unclassifiable regions, optimal DDAG based on heuristic information is proposed. Two heuristic measures are designed: static heuristic information (SHI) and dynamic heuristic information (DHL). Based on these two heuristic measures, two strategies of constructing DDAG are proposed. Experimental results based on three benchmark datasets demonstrate the superiority of our two-stage SVMs over traditional pairwise classification and DDAG based SVMs