Research On Online Learning Of Support Tensor Machine

With the rapid development of computer technology and modern Internet information,data has become more complex, and its timeliness requirements are higher. Then, the relatedproblems such as machine learning, pattern recognition and image processing have facedwith new challenges. At the same time, the online learning classification algorithm withadaptive ability would have a broad application prospect. Meanwhile, tensor, which canmaintain the natural structure and correlation in the original data, has been widely used, andsupport tensor machine (STM) has become a popular topic.Based on the stochastic gradient descent method, this paper proposes online supporttensor machine (OSTM) algorithm for tensor classification. In OSTM, its input patterns aretensors which are collected one by one in a sequence. Tensor is the natural representation ofhigh-dimensional data, keeping the natural structure information and the relationshipbetween the original data. Support tensor machine extends vector-based learning methods totensor, using tensor as input, based on the support vector machine. At last STM gains aclassification hyper plane in tensor space for data classification and recognition.Online support tensor machine (OSTM) algorithm, based on alternating projectionsupport tensor machine model and multi linear algebra, obtain a new optimization model fortensor space. Secondly, OSTM algorithm employs stochastic gradient descent method andgives out the online learning update rules based on lagrangian multiplier, in order to ensurethe model automatically adjusting with the real-time changing data. Finally, OSTMalgorithm applies tensor rank-one decomposition to replace the original tensor and assisttensor inner computation, maintaining the original information of data and greatly reducingmemory space and training time.The experiments on thirteen tensor datasets show that: compared with the onlinesupport vector machine algorithm, OSTM algorithm can provide a significant improvementin training speed with comparable test accuracy. Especially for higher-order tensors, thesuperiority of OSTM is more obvious.