Research On Multiview Induced Kernel Ensemble Regression Method

Kernel regression method is an effective non-parametric regression estimation method popular in recent years.Its performance mainly depends on whether the appropriate kernel function is selected.However,it is difficult to determine the optimal kernel function in practice,which is usually determined by cross validation of data.Based on the above problems,this study uses the ensemble regression method to find the best kernel combination representation and its parameters in multiple Reproducing Kernel Hilbert Spaces,and proposes a kernel ensemble regression method based on multi-view induction to solve the problem that it is difficult to determine the most appropriate kernel function and its parameters in the kernel regression method.The specific work is as follows:(1)Aiming at the problem of over-fitting or under-fitting and difficulty in determining the optimal kernel function and its parameters when using single kernel function to train regression model,an ensemble kernel ridge regression method based on multi-view is proposed.The main idea is to perform regression modeling operations on these virtual multi-view representations by assuming that the original data samples have multi-view representations.Finally,the multi-view linear ridge regression model is transformed into an ensemble kernel ridge regression model.Because the model is actually an ensemble regression model with multi-kernel representation,it overcomes the defect of low performance of single kernel model and has better stability and higher accuracy.The experimental results on image and UCI datasets show that the ensemble kernel ridge regression method proposed in this study has higher accuracy than other regression and ensemble methods.(2)On the basis of the above discussion,in order to improve the robustness and performance of the kernel ensemble regression model,the thesis further proposes a shared parameter research method based on multiple kernel spatial structure.The main idea is to learn the shared parameters and specific parameters of the combined kernel function in the multiple Reproducing Kernel Hilbert Spaces inspired by the consistency of the multi-view data.Among them,shared parameters are used to learn the common structure of multiple kernel representations,and specific parameters are used to learn the specific structure of the kernel.Finally,the consistency and difference of the combined kernel function can be found through parameter learning.The proposed ensemble kernel ridge regression model has the advantages of both local kernel function and global kernel function,and can adapt to different data feature sets.The experimental results on UCI datasets and character datasets further show that the shared parameter research method proposed in this study is more robust.(3)Based on the above research and multi-view induction,this thesis designs and implements a prototype system of kernel ensemble regression and classification.The system adopts the current mainstream technology and framework,uses the kernel ensemble regression and classification algorithm model studied in this thesis,and realizes the function of data regression and classification with a friendly and interactive interface,which fully verifies the superiority of the kernel ensemble regression algorithm proposed in this thesis.