Efficiently Training Ball Vector Machine in Online Way

In many real classification problems, labeled samples may be received step by step. Correspondingly, how to efficiently and progressively train classifiers should be seriously concerned. Ball vector machine (BVM), compared with support vector machine and core vector machine, can be more efficiently trained with approximate performance. In terms of the intrinsic features of BVM, an online BVM (OBVM) is presented in this paper, whose effectiveness is analytically trustable. OBVM transforms a binary classification problem into two single classification problems, in which every class is modeled by a hyper-center. Two hyper-centers are incrementally updated by using the same strategy presented in BVM. Moreover, the perpendicular bisector of plane of two hyper-centers is used to classify data. The experimental results on several standard classification datasets and a practical application indicate that, OBVM has remarkable advantages on the time complexity of training one samples per time, and also can gain high classification accuracy. Therefore, OBVM provides a new generally effective solution to train classifiers with online way and could be applied to many practical problems.