Towards Addressing an Open Problem in Coupled Matrix Tensor Factorization for Satellite Imagery Data Using Human-in-Loop

Tensor completion is a key tensor factorization task with many real-world applications. It involves predicting the missing values in a given incomplete tensor. Coupled Matrix Tensor Factorization (CMTF) is a widely used method for tensor completion. However, its widespread societal adoption has been hindered by an open problem of deriving the coupled matrix for a given incomplete tensor. When confronted with this problem, researchers tried solving it for the datasets having auxiliary information. This paper addresses the open problem in CMTF for (satellite) imagery data, where auxiliary information is unavailable. A novel model, Coupled Matrix Tensor Factorization for Satellite Images (CMTF4SI), has been proposed in this paper for tensor competition. Using the ‘Human-in-Loop’ (HiL) concept, the proposed model builds a coupled matrix by accepting the domain expertise as input and integrates it into an objective function by treating a tensor’s missing and non-missing values separately. This specific treatment is achieved by introducing the concept of an inverse mask. CMTF4SI also employs a tuning parameter (HiL-tuner) to control the influence of the human input. Experimental results on satellite imagery datasets show that integrating HiL helps improve accuracy over baseline approaches.