Flexible High-Dimensional Unsupervised Learning with Missing Data.

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• Mixture of Factor Analyzers (MFA) is a key unsupervised learning technique for high-dimensional data.
• Extensions to non-Gaussian mixtures address cluster asymmetry and heavy tails.
• Missing data presents a significant challenge in practical data analysis.

Purpose of the Study:

• To generalize the MGHFA model for handling missing values.
• To develop an efficient algorithm for parameter estimation and missing data imputation.
• To validate the proposed methodology on simulated and real datasets.

Main Methods:

• A generalized MGHFA model accommodating missing values under a missing-at-random mechanism.
• A computationally efficient alternating expectation conditional maximization (AECM) algorithm for parameter estimation.
• Investigation into missing value imputation strategies within the incomplete-data MGHFA framework.

Main Results:

• The proposed AECM algorithm effectively estimates parameters for the MGHFA model with various missing data patterns.
• The methodology demonstrates successful imputation of missing values in incomplete datasets.
• Performance evaluation shows the efficacy of the approach on both simulated and real-world data.

Conclusions:

• The generalized MGHFA model provides a robust framework for analyzing high-dimensional data with missing values.
• The developed AECM algorithm offers an efficient solution for parameter estimation and imputation.
• This work enhances the applicability of advanced mixture models in complex data scenarios.