Research Of Support Vector Machine Based On Multiple Kernel Learning

In the 1960s,Vapnik et al.put forward statistical learning theory.Based on this theory,a new learning method--support vector machine was proposed in the 1990 s.The rem arkable advantage of this method is that it effectively overcomes the drawbacks of traditional machine learning such as over-fitting,dimension disaster and local minimization,which is attributed to the rule of structural ris k minimization.Support vector machine still has good generalization ability in the case of small samples,so it has received extensive attention.With the developing and applying of support vector machines,however,there are some limitations has gradually been revealed.First,support vector m achine is sensitive to outliers and noise points.In order to solve this problem,Lin et al.proposed the concept of fuzzy support vector machine,which introduced the fuzzy membership of the sample into the support vector machine.Fuzzy support vector m achine reduces the influence of noise po ints and outliers on the final decision function to a certain extent,and it is of more anti-noise ability than traditional support vector machine.Second,it has a crucial impact on support vector m achine to select kernel function and kernel parameters.However,there are no general methods to complete this work.Multiple kernel learning,th e combination of multiple kernels,which has becom e a research hot topic in the field of machine learning in recent years.multiple kernels often can more fully character the similarity between data,especially the similarity between complex data.Therefore,the combination of multiple kernels can describe the similarity of the data more accurately and can avoid the problem of kernel selection.As an important achievement of the kernel method,multiple kernel learning can avoid the problem of kernel function selection.In order to solve prob lem of kernel selection and sensitivity to noise,we put forward multiple kernel fuzzy support vector machine and multiple kernel support vector regression,which though introduce m ultiple kernel learn ing into support vector m achines.The fuzzy support vector m achine based on multip le kernel learning significan tly improve algorithm efficiency while maintaining the anti-noise ability of the fuzzy support vector m achine.It has important theoretical and practical significance.Multiple kernel learning is seldom used in support vector regression at present.In order to solve the problem of kernel function selection in support vector reg ression,the multiple kernel learnin g based on kernel alig nment is introduced into support vector regression.The main work and achievements of the text are as follows:1.A multiple kernel learning method based on direct addition of multiple kernel functions is combined with a fuzzy support vector machine model,and we propose a fuzzy multiple kernel support vector machine based on fuzzy rough set membership.Experiments on UCI data show that the proposed fuzzy m ultiple kernel support vector m achine has better performance in prediction accuracy and ef ficiency than the classical support vector m achine,fuzzy support vector machine and multiple kernel support vector machine.2.We put forward a multiple kernel fuzzy support vector machine based on Kernel-Target alignment.Multiple kernel learning based on Kernel-Target alignment is introduced into fuzzy support vector machine.Kernel weights are calculated by m aximizing the similarity between the combined kernel and the ideal kernel.The combined kernel can describe the similarity of data more accurately.The experimental results on UCI data verify the effectiveness of the proposed method.3.Multiple kernel learning based on Kernel-Target alignment is introd uced into the support vector regression,and a m ultiple kernel support vector regression based on kernel alignment is proposed.The experimental results show that the prediction accuracy of multiple kernel support vector regression machine can keep the same or better prediction accuracy.In terms of computational cost,the multiple kernel support vector regression is much better than the traditional support vector regression.