A Modified Multi-output Support Vector Regression Machine Based on Data Dependent Kernel Function

The multi-output support vector regression machine (MSVR) is a kind of support vector regression machine which defined to deal with multi-input and multi-output problems. However, it is often difficult to improve the prediction accuracy due to its complex model structure. This paper presents a new algorithm of data dependent kernel function for MSVR, which could effectively improve the prediction accuracy of MSVR. The data dependent kernel function is constructed by using Riemannian metric. This kernel function can solve the prediction accuracy problem with complex model structure. Data dependent kernel function is also proved to satisfy the Mercer conditions in this paper. The modified MSVR is used to predict some key microbial parameters in the microbial fermentation process. The soft measuring experimental results indicate that the method is efficient to improve the prediction precision of MSVR.