Transfer Learning Research For Tensor Data

With the rapid development of information technology, information data appears a lot in daily life. How to get the useful knowledge from these information efficiently becomes more and more important. In the process of traditional machine learning, we always need to make an assumption that the test data and training data should obey the same distribution, and we can only obtain better classifier with a lot of training data. However, it is difficult to obtain large-scale sample with label in some emerging fields. If we relabeled the data samples according to the traditional way, some problems like high cost occurred. Therefore, how to overcome these shortcomings of traditional machine learning is the focus of transfer learning.Based on the existing research work, the thesis put forwards the Transfer Learning-Support Tensor Machine (TL-STM) in tensor space. Support Tensor Machine is an extension of Support Vector Machine in higher-order space. STM can classify and recognize the data with a classification hyper-plane which is acquired from tensor space. As well as traditional learning, STM can’t obtain reliable classification model in domain that lacks of training data sets. However, TL-STM algorithm can obtain a better classifier by transferring knowledge of relevant domains in the tensor space that lacks of large-scale training data. In this thesis, we propose the method to solve transfer learning for the tensor data. The detailed researches are as below:(1) The thesis explored the relationship between hyper-plane and parameters of SVM. When SVM algorithm is dealing with a high-order data sample, the defects will show out. That will lead to Support Tensor Machine algorithm directly. And through the exploration of internal relationship between parameters, we can conclude that the parameters of STM model are related, and it is efficient to do model solution with the method of alternative projection.(2) The thesis studied transfer learning of STM in second-order tensor space. When we train few labeled samples, the classification model of target domain can be obtained by combining with hyper-plane transferring knowledge in source domain. In the process of solution, model can be transformed into a solution which is to solve a series of quadratic convex programming problem by adopting computer system of alternating projection. By calculating convergence condition, we can judge if the function is convergence or not, then finish the model solution. At the end, we also analyzed and verified the feasibility of algorithm. (3) The thesis promoted the TL-STM from second-order tensor space to high-order. With the application of m-model product in tensor learning, TL-STM model can be transformed into a solution of convex optimization problem about normal vector of hyper-plane. By solving quadratic programming problem about normal vector, we can judge the convergence condition; thereafter, we can obtain TL-STM in high order tensor space.In this paper, we analyzed experiment of the algorithm in Matlab and Studio Visual software platform. Comparing with STM in classification performance, the results show that:in the target domain that lacks of training samples, TL-STM can obtain a better classifier with combination of the knowledge of hyper-plane in source domain.