Mean Decision Rule Method for Consrtucting Nonlinear Boundaries in Large Binary SVM Problems
Alexandra Makarova, Mikhail Yu. Kurbakov, Valentina V. Sulimova · 2020
In the previous work, we proposed the Mean Decision Rule method for fast approximate solution of the two- class SVM problem in a feature space. This paper generalized the proposed approach to the case of training in a space generated by a kernel function (Kernel-based Mean Decision Rule method), which allows to construct nonlinear boundaries separating objects of two classes. Experiments show that even for large training sets Kernel-based MDR, like the initial MDR method, allows to quickly find a solution that does not differ much from the exact one. Besides it is economical in memory and has a high degree of data parallelism and so it can be effectively implemented using parallel and distributed computing technologies.