A Fuzzy Support Tensor Machines based on Support Vector Data Description

Most of the traditional machine learning algorithms are based on the vector, but in tensor space, Tensor learning is helpful to overcome the over-fitting problem than vector learning. In the meanwhile, these algorithms based on tensor require a smaller set of decision variables as compared to those approaches based on vector. Support tensor machine (STM) is a prevalent machine learning approach, and STM is widely applied to practical many problems. However, STM is sensitive to noise or outliers in the sample dataset. Fuzzy support tensor machine (FSTM) can partly overcome this shortcoming by assigning different fuzzy membership to different training samples. How to choose the appropriate fuzzy memberships is one of the important things for machine learning. At present, most of people define a fuzzy membership by introducing the distance between each point and its class centre. In this paper, We present a novel methods to compute fuzzy memberships based on support vector data description (SVDD), our experimental results on ORL database and Yale database show that the improved method can reduce the effect of outliers and improve the rate of classification accuracy.