Model-based statistical depth for matrix data

Guanyu Hu (Center for Spatial Temporal Modeling for Applications in Population Sciences, Department of Biostatistics and Data Science, University of Texas Health Science Center, Houston, Tx., U.S.A.)

Wei Wu (Department of Statistics, Florida State University, Tallahassee, Fl., U.S.A.)

Abstract

The field of matrix data learning has witnessed significant advancements in recent years, encompassing diverse datasets such as medical images, social networks, and personalized recommendation systems. These advancements have found widespread application in various domains, including medicine, biology, public health, engineering, finance, economics, sports analytics, and environmental sciences. While extensive research has been conducted on estimation, inference, prediction, and computation for matrix data, the ranking problem has not received adequate attention. Statistical depth, a measure providing a centeroutward rank for different data types, has been introduced in the past few decades. However, its exploration has been limited due to the complexity of the second and higher orderstatistics. In this paper, we propose an approach to rank matrix data by employing a model-based depth framework. Our methodology involves estimating the eigen-decomposition of a 4th-order covariance tensor. To enable this process using conventional matrix operations, we specify the tensor product operator between matrices and 4th-order tensors. Furthermore, we introduce a Kronecker product form on the covariance to enhance the robustness and efficiency of the estimation process, effectively reducing the number of parameters in the model. Based on this new framework, we develop an efficient algorithm to estimate the model-based statistical depth. To validate the effectiveness of our proposed method, we conduct simulations and apply it to two real-world applications: field goal attempts of NBA players and global temperature anomalies.

Keywords

matrix data, data depth, covariance tensor, eigen-decomposition

The full text of this article is unavailable through your IP address: 3.144.3.235

Dr. Hu’s research is partially supported by NSF awards SES-2243058 and DMS-2412923.

Received 1 March 2023

Accepted 12 December 2023

Published 1 February 2024