Bayesian Tensor Tucker Completion With a Flexible Core
Xueke Tong, Lei Cheng, Yik‐Chung Wu
Abstract
Tensor completion is a vital task in multidimensional signal processing and machine learning. To recover the missing data in a tensor, various low-rank structures of a tensor can be assumed, and Tucker format is a popular choice. However, the promising capability of Tucker completion is realized only when we can determine a suitable multilinear rank, which controls the model complexity and thus is essential to avoid overfitting/underfitting. Rather than exhaustively searching the best multilinear rank, which is computationally inefficient, recent advances have proposed a Bayesian way to learn the multilinear rank from training data automatically. However, in prior arts, only a single parameter is dedicated to learn the variance of the core tensor elements. This rigid assumption restricts the modeling capabilities of existing methods in real-world data, where the core tensor elements may have a wide range of variances. To have a flexible core tensor while still retaining succinct Bayesian modeling, we first bridge the tensor Tucker decomposition to the canonical polyadic decomposition (CPD) with low-rank factor matrices, and then propose a novel Bayesian modeling based on the Gaussian-inverse Wishart prior. Inference algorithm is further derived under the variational inference framework. Extensive numerical studies on synthetic data and real-world datasets demonstrate the significantly improved performance of the proposed algorithm in terms of multilinear rank learning and missing data recovery.