Research On Robust Non-parallel Plane Classifier For Multi-model Data

Support Vector Machine(SVM),as one of the very popular classification tools in the fields of pattern recognition and data mining,which was born in 1960 s and has been widely used in the past decades.The extended version of SVM: Generalized Eigenvalue Proximal Support Vector Machine(GEPSVM),improves classification performance while reducing calculation complexity.And with the rapid growth of information technology,the popularity of multimodal data has spawned the emergence and development of multi-view learning(MVL).Compared with the traditional single-view method,multi-view learning could excavate the complementary and consistent information among different views to improve learning performance.Multi-view leaning with Generalized Eigenvalue Proximal SVM(MvGSVM),which cleverly merges multiview learning with GEPSVM,by using multi-view co-regularization to find consistency between different views.However,for most multi-view methods,the optimization of the target model is based on the square L2-norm distance measures,which is susceptible to outliers and noise.Therefore,three robust multi-view classification algorithms based on MvGSVM are proposed in this paper to discuss and solve the problem of insufficient robustness.The main research contents of this article are as follows:1.Based on the good classification performance of MvGSVM,this thesis proposes a new method called Multi-view learning with robust GEPSVM(Lp,s-MvGEPSVM).By introducing Lp-norm minimization and Ls-norm maximization instead of L2-norm,the distances from the hyperplane to the positive and negative samples are calculated respectively.In order to solve the non-convex optimization problem based on Lp-norm,an effective iterative solution algorithm is designed,and the convergence of the algorithm is proved theoretically.A large number of experiments on the face dataset and UCI dataset have showed the advantages of the new method in classification accuracy,robustness and flexibility.2.This thesis further proposes a new multi-view learning method: multi-view learing with robust double-sided twin SVM(MvRDTSVM),expanding the MvGSVM to solve QP type problems and using L1-norm distance metric to improve the robustness.Moreover,MvRDTSVM synergistically considers the double-sided strategy,which relaxs the constraint by allowing negative class samples to be distributed on both side of the positive class fitting hyperplane.In the face of more complex optimization goals,a new iterative algorithm is designed and the proof of convergence is given.Experiments on UCI datasets shows the effectiveness and feasibility of the proposed algorithm.3.MvRDTSVM needs to solve a series of QP problems,which will undoubtedly bring greater computational cost.Since,according to the idea of least squares,this thesis replaces the inequality constraints with equality constraints,and propose a fast multi-view learning with robust 2-sided twin SVM(MvFRDTSVM).By solving a system of linear equations in each iteration,the calculation speed is greatly accelerated.The proof of convergence guarantees the feasibility of the algorithm theoretically.Similarly,the experimental results on UCI datasets show the advantages of the new method being fast and efficient.