Listen: Adaptive Low-Separation-Rank Ridge Regression For Tensor Covariates
Abstract
Modern applications increasingly generate tensor-structured data (i.e., multiway arrays), which are often used as covariates in regression problems. In this work, we focus on ridge regression with tensor covariates, where a common approach is to vectorize each tensor and estimate an unstructured coefficient vector. However, this approach fails to exploit the tensor structure of the regression coefficients and can create high-dimensional, ill-posed estimation problems when sample sizes are limited. Prior work has addressed these challenges by imposing low-separation-rank (LSR) structure on the coefficient tensor, thereby reducing the effective complexity of the model. However, existing factorized LSR methods require the separation rank to be specified a priori, making model selection computationally expensive and the resulting estimator sensitive to rank misspecification. We propose LISTEN, an implicitly regularized LSR-structured ridge regression framework that enforces structure via a convex regularizer rather than an explicit rank-constrained factorization. We develop an alternating minimization algorithm with ADMM-based updates and establish a stationarity guarantee under exact block minimization. Synthetic experiments show that LIS-TEN adapts as the coefficient tensor deviates from a compact LSR structure, while neuroimaging experiments demonstrate competitive predictive performance relative to the considered baselines.