Keywords: Tensor, coclustering, alternating least squares, generalized linear model
Multi-way tensor array data are frequently encountered in real applications. Oftentimes, it is of interest to find coherent coclusters consisting of subsets of features in each mode within a tensor. Coclusters identify feature groups in different modes simultaneously and carry important interpretation. In this talk, we will introduce a new method for identifying coclusters based on a regularized alternating least squares approach. We further generalize the method to deal with multi-way tensors with non-Gaussian entries. We will demonstrate the superior performance of the method using simulations and a range of real applications.