| With the rapid development of cloud computing, networking and social networks, the advent of big data is coming. The problem is no longer how to collect data but mine big data. Considering that the input patterns are usually high-order tensors in the fields of machine learning, pattern recognition, data mining, image processing and computer vision. If tensor objects are reshaped into vectors, this direct reshaping breaks the natural structure and correlation in the original data. Therefore, a good tensor learning method is emerged to be proposed. Recently, some researchers present a tensor factorization based least squares support tensor machine(TFLS-STM) which combines the merits of supervised tensor learning framework and tensor rank-one decomposition.Nowadays, tensor objects such as sensor data, images and video data are sequentially learned in many practical applications. Considering that the amount of learning data is unknown, we have no idea to store and batch learning all of them. In this study, we present an online sparse learning algorithm—online least squares support tensor machine(OLS-STM). We test whether new tensor object is approximately linearly dependent on the dictionary tensors which are approximately linearly independent with each other in the dictionary. We eliminate the redundant object and only use the objects in the dictionary to update the learning model. The computational complexity of new algorithm is independent of the number of objects.We have designed a series of comparative experiments on 13 tensor datasets to verify the effectiveness of OLS-STM. We compare this new algorithm with principal component analysis based least squares support vector machine(PCALS-SVM) and tensor factorization based least squares support tensor machine(TFLS-STM) on training time and precision. The results show that OLS-STM enjoys faster training speed and has low computational complexity on datasets which has a large number of samples and redundant samples. |
PDF Full Text Request