Study On One Class Support Tensor Machine And Its Algorithm

With the advent of the era of big data, vector is no longer the only data type in data mining, there are various types of data that need to be mining. For instance, Fingerprint recognition, face recognition, text classification and so many other application problems need to process tensor data rather than vector data. The traditional method of processing tensor data is converted it into vector data then use vector based algorithm to process the vector data. However, this approach ignores the structural of the tensor data itself, and the high dimension of the vector which converted by tensor will cost a very long time for the algorithm to calculate. So the researching of tensor based algorithm that can efficiently exploit the tensor data is very necessary.One Class Support Vector Machine(OC-SVM) is an algorithm inspired by Support Vector Machine(SVM). The difference between OC-SVM and SVM is that OC-SVM is mainly used in anomaly detection problems and SVM is mainly used in classification problems. This article is focus on One Class Support Tensor Machine(OC-STM), an anomaly detection algorithm make use of tensor kernels and tensor decomposition to process tensor data. Studies have shown that Support Tensor Machine(STM) which make use of tensor decomposition and tensor kernels to process tensor data can improve the classification accuracy and decrease the training time. Theoretically CP decomposition can reduce the dimension of the tensor data, so it can decrease the training time; tensor kernel function can make use of the architecture of the tensor data, so it can improve the accuracy. Base on the above theory and the former experiment results, this method could probably also improve the precision of other support vector based models and decrease the training time. Finally we will experiment on two order tensor database Yale-B, ORL, CMU PIE and three order tensor database USF HumanID. The experiment will show the accuracy and training time of OC-STM, and we will compare its result with OC-SVM.