Abstract
Real-world datasets often contain missing or corrupted values. Completingmultidimensional tensor-structured data with missing entries is essential fornumerous applications. Smoothness-constrained low-rank factorization modelshave shown superior performance with reduced computational costs. Whileeffective at capturing global and long-range correlations, these modelsstruggle to reproduce short-scale, high-frequency variations in the data. Inthis paper, we introduce the Generalized Least Squares Kernelized TensorFactorization (GLSKF) framework for tensor completion. GLSKF integratessmoothness-constrained low-rank factorization with a locally correlatedresidual process; the resulting additive structure can effectively characterizeboth global dependencies and local variations. In particular, we define thecovariance norm to enforce the smoothness of factor matrices in the globallow-rank factorization, and use structured covariance/kernel functions to modelthe local processes. For model estimation, we develop an alternating leastsquares (ALS) procedure with closed-form solutions for each subproblem. Toefficiently handle missing data, GLSKF utilizes projection matrices thatpreserve the Kronecker structure of covariances, facilitating fast computationsthrough conjugate gradient (CG) and preconditioned conjugate gradient (PCG)algorithms. The proposed framework is evaluated on four real-world datasetsacross diverse tasks: traffic speed imputation, color image inpainting, videocompletion, and MRI image reconstruction. Experimental results confirm thatGLSKF delivers superior effectiveness and scalability, establishing it as arobust solution for multidimensional tensor completion.