Performance Improvement Based On Pinball Support Vector Machine

For classification problems,the pinball support vector machine based on the quantile distance of the data has better robustness than the classic hinge loss support vector machine based on the shortest distance of the data.This paper is based on the pinball support vector machine method,combined with the ”distributed” and ”composite” ideas,respectively,to improve the performance of methods and algorithms from both computational efficiency and parameter estimation effectiveness.In order to solve the problems of computer memory limitation and computational cost caused by modern large-scale data,this paper proposes and studies a pinball loss linear support vector machine(pin-SVM)distributed estimation method suitable for binary classification tasks in a big data environment.In this paper,based on the smoothing approximation of the pinball loss function,we iteratively solve the parameters and obtain the multi-round distributed linear estimation of the parameters of the pin-SVM.Through numerical simulation and its application on real data sets,it can be verified that our method not only has high accuracy for parameter estimation,but also has robustness to noise.At the same time,distributed computing greatly accelerates the calculation speed and reduces the calculation cost.In order to further improve the parameter estimation effectiveness of pin-SVM,we propose a composite pinball support vector machine method.Firstly,we give the linear form of the composite pinball support vector machine,and theoretically prove the asymptotic convergence properties of the linear composite pinball support vector machine.Secondly,we give the nonlinear kernel form of the composite pinball support vector machine,and derive its dual problem from this form.Finally,for the practicability of the proposed method,we design a sequential minimal optimization algorithm to solve the optimization problem of composite pinball support vector machine,and show that our method has excellent performance in the effectiveness of parameter estimation through numerical experiments.