An ensemble approach to tensor learning

Abstract

Motivated by recent development of tensor regression modeling, we propose a novel tensor ensemble learning (TEL) approach. While CANDECOMP/PARAFAC (CP) decomposition is an efficient technique to reduce the number of parameters in tensor covariate, determining an appropriate CP rank remains uncertain in practice. Furthermore, the structural complexity of tensor often varies across different spatial regions. To explore the intrinsic tensor structure, we design different tensor partition strategies and divide tensor into disjoint blocks to form candidate models. A model ensemble method is then developed to explore the uncertainty in both tensor block structure and CP rank. To assign weights to candidate models and enhance predictive performance, we leverage the predictability, computability, and stability framework for veridical data science. Simulation studies are carried out to assess the performance of TEL under varying levels of tensor complexity. We further apply TEL to two real data applications about glaucoma management using fundus image and cognitive ability prediction in Alzheimer’s disease using neuroimaging. Our numerical studies illustrate the superiority of TEL over existing methods and highlight its promise for complex tensor data analysis.