Title: Convex tensor clustering with applications to online advertising
Authors: Brian Gaines - North Carolina State University (United States)
Eric Chi - North Carolina State University (United States)
Will Wei Sun - University of Miami School of Business Administration (United States) [presenting]
Hua Zhou - UCLA (United States)
Abstract: Tensors, as a multi-dimensional generalization of matrices, have received increasing attention in industry due to its success in modeling data with complex structures. One typical circumstance is in online advertising, where user click behavior on different ads from multiple publisher platforms forms a user-ad-publisher tensor. The goal is to simultaneously group users, ads, and publishers for better targeted advertising. We will discuss a convex formulation of the tensor clustering problem, which is guaranteed to obtain a unique global minimum. It generates an entire solution path of clusters in all tensor modes governed by one tuning parameter, and thus alleviates the need to specify the number of clusters a priori. The finite sample error bound of the proposed estimator reveals an interesting bless of dimensionality phenomenon in the tensor clustering. To demonstrate the potential business impact of our method, we conduct convex clustering on the user-ad-publisher tensor data obtained from a major online company. Our clustering results provide interesting insights in understanding the user click behavior.