Tensor Subspace Cluster

As a typical unsupervised learning technique, subspace clustering learns the subspaces of data and assigns data into their respective subspaces, which is important for a number of data processing applications. Traditional subspace clustering is based on matrix computation, and it is inevitable to lose some structural information when dealing with multidimensional data. To alleviate performance degeneration from matricization or vectorization, tensor subspace clustering is used to learn directly in tensor subspace. In this chapter, we mainly introduce two tensor-based cluster models, including K-means and self-representation, respectively. In particular, different subspace clustering models based on different tensor decompositions and corresponding algorithms are outlined in details, such as Tucker decomposition and t-SVD. To demonstrate the performance in practical applications, we apply tensor subspace clustering in clustering for heterogeneous information networks, multichannel ECG signal clustering, and multi-view data clustering. Experimental results show the tensor subspace clustering has superior performance than its matrix counterpart.