Research On Fast Online Support Tensor Machine Classification Algorithm

With the rapid development of information technology, all trades and professions have accumulated large amounts of data, which most to be expressed in tensor can preserve the structure information of the original data. Besides, facing the big data era, timeliness requirements are also increasing. As a result, the study of online classification algorithms for tensor data is very meaningful.Based on fast online classification algorithm(Online LASVM) with vector input, this article mainly focus on online classification algorithm with tensor input. The main contents are as follows:(1) We promote Online LASVM into tensor space, and propose fast online support tensor machine classification algorithm(FOSTM), which is aimed at solving the online classification problem of obtaining sequence tensor data. The algorithm is first based on a linear support higher-order tensor machine, then by virtue of tensor CP decomposition technique, we construct an effective tensor kernel function which can keep the data structure of the original information, and obtain nonlinear support tensor machine learning model. Finally to solve the optimization model, then obtain FOSTM which with tensor data as the input pattern.(2) By improving the reduction process in FOSTM algorithm, we propose fast online support tensor machine classification algorithm based on duality gap(FOSTM-G). Whether Online LASVM, or its promotion FOSTM perform an incremental learning, then follow perform reduction process, this will prone to owe optimization middle model, in turn affect the performance of online classifier. Therefore, on the basis of FOSTM, we introduce the idea of minimizing the gap between the primal and the dual to improve the reduction process.(3) Experiments on 12 tensor datasets verify the effectiveness of the FOSTM and FOSTM- G, and show that: compared with Online LASVM, FOSTM and FOSTM- G have significantly improvement in learning efficiency with compared test accuracy, especially for higher-order tensors, the superiority of the two online algorithms under tensor space is more obvious. In addition, for all data sets, the FOSTM – G's test accuracy is higher or equal to FOSTM algorithm, but training time greatly reduced.