Michigan Technological University
HYBRID MIXTURES OF FACTOR ANALYZERS FOR HIGH DIMENSIONAL DATA
Abstract
dc:description.abstractFactor analysis is a powerful tool for modeling latent structures in high-dimensional data, traditional approaches assume a single global structure, limiting their ability to capture heterogeneity. The Mixture of Factor Analyzers (MFA) extends classical factor analysis by modeling data as a mixture of Gaussian-distributed local subspaces, effectively uncovering cluster-specific latent structures. However, MFA relies on Gaussian mixtures, making it sensitive to outliers and ill-suited for heavy-tailed data. The Mixture of $t$-Factor Analyzers (M$t$FA) addresses these limitations by incorporating multivariate $t$-distributions, improving robustness. Despite their advantages, both MFA and M$t$FA face significant computational challenges in high-dimensional settings, particularly due to costly covariance matrix operations and slow convergence of the Expectation-Maximization (EM) algorithm. In this collection of work, we propose a hybrid approach that integrates a matrix-free algorithm within the EM framework to improve computational efficiency for both MFA and M$t$FA. Our methods preserve the full covariance structure while leveraging matrix-free strategies to optimize computations in high dimensions. This approach maintains the interpretability and flexibility of mixture-based factor models while making them more scalable. Empirical evaluations on synthetic and real-world data reveals that our method significantly improves speed, robustness, while preserving clustering accuracy compared to conventional MFA and M$t$FA models driven by standard EM algorithms. These findings highlight the effectiveness of matrix-free strategies in advancing mixture-based latent variable models for high-dimensional data analysis.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy in Statistics (PhD)
- Level thesis:degree_level
- Open Access Dissertation
- Discipline thesis:degree_discipline
- Department of Mathematical Sciences
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kareem, Kazeem Abiodun
- Committee members dc:contributor.committeemember
-
- Qiuying Sha
- Kui Zhang
- Nathir A. Rawashdeh
- Contributors dc:contributor
-
- Fan Dai
Subjects
dc:subject × 13Rights
- Licence dc:rights.license
Identifiers
dc:identifier.*- Identifier
- https://digitalcommons.mtu.edu/etdr/1901