Reweighted Robust Support Vector Regression Method

A reweighted robust support vector machine (WRSVR) is presented for regression estimation and function approximation by soft pruning of outliers. The essential idea of the approach is to reweight regularization parameters to softly prune the outliers. Fundamentally, the regression function is improved iteratively. First, an approximate regression function is obtained using support vector regression (SVR), and then a weighted SVR objective function is generated and a new approximate regression function is got by solving the corresponding minimization problem. Based on the new approximate regression function, another weighted SVR objective function is found and a new approximate regression function is obtained. This procedure is repeated until it converges. The new approach is conceptually simple, insensitive to outliers and easy to implement. Numerical simulations show that it is more robust than the currently used Standard SVR, Robust Support Vector Regression Networks and Weighted Least Square SVM.