Research On Multidimensional Output Support Vector Regression And Application

Support vector regression (SVR) is a machine learning method developed on the basis ofstatistical learning theory and optimization theory, it has inherited the support vector machine inovercoming the " curse of dimensionality " and " over-learning " etc.It has an intuitive geometricalinterpretation and a perfect mathematical form, thus promoting the development of predictiveregression.This paper studies the multidimensional output support vector regression and its applicationproblems. The main innovation is as follows:1. Proposed multidimensional output twin support vector regression(MTSVR). With theexpansion of the kernel mapping space to replace the original one-dimensional output kernel space,twin support vector regression (TSVR) has been improved to adapt to the multidimensional outputof the algorithm. Experimental results show that in the case of two-dimensional output, MTSVRalgorithm acquire higher accuracy than TSVR algorithm.2. Proposed a multidimensional output support vector regression based on quadratic lossfunction (L2-MSVR). Two norm loss function is introduced to the standard support vectorregression (SVR), then applies to multidimensional output situation. Experimental results showthat in the output of two-dimensional case, L2-MSVR has higher regress accuracy, good stabilityand faster.3. Proposed multidimensional output support vector regression based on the expansion of thekernel function and quadratic loss function (L2-MkSVR). In general, combine the MTSVR’s ideawith the L2-MSVR’s idea, applied to a standard support vector regression, the experimentalstructure shows that in the case of certain parameters, the L2-MkSVR algorithm is more stable andprediction accuracy is higher than the L2-MSVR algorithm’s.