A support vector machine with maximal information coefficient weighted kernel functions for regression

Abstract

For some time past, support vector machine (SVM) has been generally used in pattern recognition, classification and prediction. However, in traditional SVM arithmetic, various kernels cannot recognize the importance of the feature vector properties, so the prediction accuracy seems to be unsatisfactory. For purpose of optimizing the problem, this paper proposes an modified support vector regression(SVR) method with maximal information coefficient(MIC) weighted kernel function(MWSVR) in this paper. After analyzing the correlation between feature variables and output variables through MIC and giving each feature weight on the radial basis function (RBF) kernel function, and then training the model and predicting the tendency. Adding MIC-weight vector on kernel function can help to identify the importance of the feature vectors and obtain better prediction accuracy. MWSVR has practicability and generalization.