Abstract:Low-Rank Tensor Completion, a method which exploits the inherent structure of tensors, has been studied extensively as an effective approach to tensor completion. Whilst such methods attained great success, none have systematically considered exploiting the numerical priors of tensor elements. Ignoring numerical priors causes loss of important information regarding the data, and therefore prevents the algorithms from reaching optimal accuracy. Despite the existence of some individual works which consider ad hoc numerical priors for specific tasks, no generalizable frameworks for incorporating numerical priors have appeared. We present the Generalized CP Decomposition Tensor Completion (GCDTC) framework, the first generalizable framework for low-rank tensor completion that takes numerical priors of the data into account. We test GCDTC by further proposing the Smooth Poisson Tensor Completion (SPTC) algorithm, an instantiation of the GCDTC framework, whose performance exceeds current state-of-the-arts by considerable margins in the task of non-negative tensor completion, exemplifying GCDTC's effectiveness. Our code is open-source.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: in Low - Rank Tensor Completion (LRTC), how to systematically utilize the Numerical Priors of tensor elements. Specifically, although the existing LRTC methods have achieved remarkable success, they usually ignore the numerical characteristics of tensor elements, which leads to the algorithms being unable to achieve optimal accuracy. Although some task - specific works have considered specific numerical priors, there is not yet a general framework to systematically integrate these numerical priors. ### Main contributions of the paper: 1. **Propose a general framework**: The author proposes the Generalized CP Decomposition Tensor Completion (GCDTC) framework, which is the first low - rank tensor completion framework that can systematically utilize numerical priors. 2. **Theoretical analysis**: The author conducts theoretical analysis and derives a method for jointly optimizing numerical priors and inter - element prior losses. 3. **Specific application**: By using the Poisson Distribution as a numerical prior, the author proposes the Smooth Poisson Tensor Completion (SPTC) algorithm and verifies its effectiveness in non - negative integer tensor completion tasks. Experimental results show that SPTC is significantly superior to the current state - of - the - art algorithms in multiple metrics. ### Mathematical formula representation: - **Objective function**: \[ \max_{A^{(1)}, A^{(2)}, \ldots, A^{(N)}} \left( \prod_{i \in \Omega} p(t_i | x_i) \right) \] To facilitate optimization, it is converted to the logarithmic form: \[ L_1 = -\ln \left( \prod_{i \in \Omega} p(t_i | x_i) \right) = -\sum_{i \in \Omega} \ln p(t_i | x_i) \] - **Inter - element prior**: \[ L_2 = \sum_{n = 1}^N L_2^{(n)} \] where \( L_2^{(n)} \) is the inter - element prior loss of the \(n\)th factor matrix. - **Overall optimization objective**: \[ \min_{A^{(1)}, A^{(2)}, \ldots, A^{(N)}} \left( -\sum_{i \in \Omega} \ln p(t_i | x_i) + \sum_{n = 1}^N L_2^{(n)} \right) \] - **Loss function under Poisson distribution**: \[ L_1(X) = \sum_{i \in \Omega} (x_i - t_i \ln x_i) \] - **Gradient update**: \[ A^{(n)} \leftarrow A^{(n)} - \alpha \left( \frac{\partial L_1}{\partial A^{(n)}} + \frac{\partial L_2^{(n)}}{\partial A^{(n)}} \right) \] ### Experimental results: The author conducts experiments on standard test images, and the results show that the performance of SPTC on multiple datasets is significantly better than that of the existing state - of - the - art algorithms. For specific results, see the PSNR values in Table 1. In most cases, SPTC obtains the highest PSNR values. ### Conclusion: This paper successfully solves the problem of how to systematically utilize numerical priors by proposing the GCDTC framework, and verifies the effectiveness of the framework in practical tasks through the SPTC algorithm. This work provides new directions and tools for research in the field of low - rank tensor completion.

A generalizable framework for low-rank tensor completion with numerical priors

A generalizable framework for low-rank tensor completion with numerical priors

Accurate Tensor Completion via Adaptive Low-Rank Representation

A dual framework for low-rank tensor completion

Beyond Low Rank: A Data-Adaptive Tensor Completion Method

Tensor Completion via Complementary Global, Local, and Nonlocal Priors

Low rank tensor completion with sparse regularization in a transformed domain

An Optimal Statistical and Computational Framework for Generalized Tensor Estimation

Mixed norm regularized models for low-rank tensor completion

Matrix factorization for low-rank tensor completion using framelet prior

Robust Low-tubal-rank Tensor Completion based on Tensor Factorization and Maximum Correntopy Criterion

Simultaneous Tensor Decomposition and Completion Using Factor Priors

Tensor Deli: Tensor Completion for Low CP-Rank Tensors via Random Sampling

Nonconvex Robust High-Order Tensor Completion Using Randomized Low-Rank Approximation

Low-CP-Rank Tensor Completion via Practical Regularization

Tensor Completion Via Collaborative Sparse and Low-Rank Transforms

Exploiting Global Low-Rank Structure and Local Sparsity Nature for Tensor Completion.

An Efficient Tensor Completion Method Via New Latent Nuclear Norm

T-product factorization based method for matrix and tensor completion problems

Provable Tensor-Train Format Tensor Completion by Riemannian Optimization