Comparison Study On Multi-category Classification With Binary SVMs

Support Vector Machine (SVM for short) is a new machine learning method basedon Statistical Learning Theory. Because SVM has the advantages of the global optimum,simple structure, strong generalization ability, it has been broadly researched andapplied. But the SVM approach was originally designed to solve binary classificationproblems, how to effectively extend SVM for multi-category classification problem isstill an on-going research issue.Based on the basic theory and basic properties of the SVM, multi-categoryclassification methods which based on binary classifiers were studied in this paper.Main research works are as follows.In this paper, common multi-category classification methods including OAA, OAO,DAG and ECOC are summarized. Then we give a summary of the principle of theclassification theory, error correction capability, and the computational complexitycomparison. Introduce a new method which via data-driven topology-preserving outputcode (TPOC for short).In this paper, we use multi-category classification algorithms, including OAA,DAG, ECOC, DECOC and TPOC, to do the comparative experiments of recognitionrate, complexity and training time on5data sets, including iris, segment, NCI, ISOLETand Letter. The results show that, DAG algorithm has high recognition ability andgeneralization ability. Compared with ECOC and DECOC, TPOC algorithm caneffectively reduce the number of support vectors under the condition of the recognitionrate less loss.