Boundary Points Based Locally Linear Support Vector Machine

The present era is also called big-data era.Its specialty is that people are surround by big-data.Against this background,machine learning is becoming a powerful tool for big-data processing.However,to process various big-data,conventional machine learning algorithms are faced with much challenge.The first is how to process big-data in an efficient and effective way.In addition,how to efficiently process online data also remains unsolved.Support Vector Machine(S VM),which is one of the most important algorithms in machine learning,is also faced with these problems.In brief,linear SVM is efficient and able to process online data but with poor performance,and kernel SVM owns powerful classification performance but is not efficient and cannot process online data.To design a SVM algorithm which is able to process online data both efficiently and effectively,we focus on the research on boundary points based locally linear SVM.We propose such a solution:partition nonlinear data into subparts based on bound-ary points and classify each part by a linear SVM.Specifically,we first propose a method called Local Linear SVM based on Boundary Anchor Points Enchoding(LL-BAP)to achieve the efficiency of linear SVM and power of kernel SVM.LLBAP uses an adaptive way to find the boundary points of training data.Then,boundary points based local coding is used to partition training data into approximately linearly separa-ble parts.Each part corresponds to a linear SVM.We also define a loss function about boundary points and linear SVMs and then use stochastic gradient descent to optimize the LLBAP model.Since LLBAP cannot process online data,we make some enhance-ments on LLBAP and propose our second method-INLEX(Incremental Network with Local EXperts Ensemble).INLEX uses an entropy based online method to find bound-ary points.Moreover,INLEX combines boundary points learning with classifying training and is able to process online data.The experimental results on large benchmark data sets demonstrate that,the pro-posed method is both training and testing efficient compared with kernel SVM;its ef-ficiency and classification accuracy also outperform other state-of-art related methods.INLEX's experimental results also demonstrate that INLEX owns stable incremental performance and better classification ability than other related methods.