Energy-based structural least squares MBSVM for classification

摘要

Multiple birth support vector machine (MBSVM) is an extension of twin support vector machine on multi-class classification problem. In MBSVM, the size of each QP problem is restricted by the number of patterns in one of the K classes, so the computational complexity of MBSVM is much lower and the training speed of it is faster than the existing multi-class SVM. However, MBSVM neglects the structural information of data which may contain some significant prior knowledge for training classifiers. In this paper, we first present an improved version of structural least square twin support vector machine (S-LSTWSVM), called energy-based structural least square twin support vector machine (ES-LSTWSVM), which converts the constraints of the S-LSTWSVM into an energy-based model by introducing an energy for each hyperplane. Then we use the strategy of “rest-versus-one” in MBSVM to extend ES-LSTWSVM into the multi-class classification, called energy-based structural least squares MBSVM (ESLS-MBSVM). In order to prove the validity of ESLS-MBSVM, the experiment has been performed on UCI datasets. The experimental results show that our ESLS-MBSVM is effective and has good classification performance. In order to better illustrate the experimental results, we use Friedman test and ROC analysis for statistical comparisons.