Research On Some Problems Of Projection Twin Support Vector Machine

Support vector machine(SVM)based on statistics theory is one of the important research methods in the field of pattern recognition and Machine learning.It has intuitive geometric interpretation and perfect mathematical form,and is widely used in the fields of image classification,speech recognition and biological information.With the development of research,inspired by the idea of SVM,Khemchandani et al.proposed the Twin support vector machine(TWSVM)in 2007.By looking for a pair of non-parallel classification hyperplanes,TWSVM makes similar samples close to one plane and as far away from the other plane as possible.Then,non-parallel hyperplane support vector machines are widely studied and many improved algorithms are proposed.Projection twin support vector machine(PTSVM)is one of the improved non-parallel hyperplane support vector machines.In view of the good classification performance of PTSVM,many scholars have conducted in-depth studies on it and achieved fruitful results.However,there are still some problems to be solved in this algorithm.In this paper,some studies have been carried out on the parameter selection and kernel function selection of PTSVM,and the specific contents are as follows.1.There are many parameters in the PTSVM,which have a great impact on the performance of the algorithm.Grid search is usually used to optimize the parameters.However,the grid search method generally runs for a long time,which makes the model inefficient.Aiming at the problem of parameter selection of PTSVM,this paper combines particle swarm optimization(PSO)algorithm with PTSVM algorithm to propose a projection twin support vector machine based on PSO algorithm,which gives full play to the global search ability of PSO algorithm to find the optimal parameters better and improve the performance of the algorithm.Experimental results show that the proposed algorithm not only improves the classification accuracy,but also reduces the total running time.2.For the nonlinear PTSVM algorithm,at first,The sample points to the high-dimensional feature space through the nonlinear mapping.The selection of the kernel function often has a great impact on the performance of the nonlinear PTSVM.The most commonly used kernel functions have their own advantages and disadvantages in learning ability and generalization performance.So,they cannot adapt to the data with uneven or irregular sample distribution.In these cases,nonlinear PTSVM using only a single kernel will show a wide range of the effects.In order to solve the problem of kernel function selection,this paper combines the above three kinds of kernel functions,constructs a new linear combination kernel function,and proposes a projection twin support vector machine based on the proposed combination kernel.The effectiveness of the proposed algorithm is verified on different data sets.