Inference of Poisson count processes using low-rank tensor data

A novel regularizer capturing the tensor rank is introduced in this paper as the key enabler for completion of three-way data arrays with missing entries. The novel regularized imputation approach induces sparsity in the factors of the tensor's PARAFAC decomposition, thus reducing its rank. The focus is on count processes which emerge in diverse applications ranging from genomics to computer and social networking. Based on Poisson count data, a maximum aposteriori (MAP) estimator is developed using the Kullback-Leibler divergence criterion. This probabilistic approach also facilitates incorporation of correlated priors regularizing the rank, while endowing the tensor imputation method with extra smoothing and prediction capabilities. Tests on simulated and real datasets corroborate the sparsifying regularization effect, and demonstrate recovery of 15% missing RNA-sequencing data with an inference error of -12dB.