The Research Of Sparse And Low-rank Subspace Clustering Method For Sequential Data

With the rapid development of network technology, a huge amount of data have been produced by the information perception and network applications, as the era of big data has arrived. At present, the vast majority of network information is data with no label, how to quickly extract valuable information which is hidden in large data and accessing to semantic information of these data is the biggest challenge for large data processing and utilization. Unsupervised clustering analysis are effective methods for access to the data labels, which is a hot research topic in data processing and analysis. Especially, because of its higher clustering performance, researchers have paid extensive attention to the subspace clustering method. In recent years, the most representative subspace clustering method is a low rank representation(LRR)and sparse subspace clustering(SSC) method, they are based on the assumption that each data can be linear represented by the data from the same subspace, which need the coefficient matrix be sparse or low rank respectively. These methods have very good performance in high-dimensional data processing and applications, such as face recognition, handwritten word and text clustering.Although low rank representation(LRR) and sparse subspace clustering(SSC)method have obtained good application effect, they ignored some data structure properties, and these properties has great influence on clustering, in the face of actual complex data. For example, many data to be clustered in practical applications have sequence characteristics, including core spectral data, video, motion sequences, etc.But the existing subspace clustering methods seldom use the sequence attributes of the data to cluster. In order to reflect and describe the structural properties of complex data, and obtain the ideal clustering effect, in this paper, we will study the subspace clustering problem for sequential data and propose a new clustering method which is verified by a lot of experiments. The main work of this paper is summarized as follows:(1) A better clustering method, ordered subspace clustering with block-diagonal prior is proposed. The representation matrix is block-diagonal and constraint by a time-space ordered penalty which is based on prior knowledge.(2) By establishing a unified model of ordered subspace clustering and 3D reconstruction of non-rigid motion, we proposed a new method named ordered subspace clustering for sequential non-rigid motion by 3D reconstruction. The structure of ordered subspace, makes the 3D reconstruction more accurate, and 3D motion reconstruction provides more information for clustering, which enhances the accuracy of the subspace segmentation. The two complement and promote each other,improved the performance of clustering.