Sparse covariance estimation in heterogeneous samples

Area of Science:

• Statistics
• Econometrics
• Machine Learning

Background:

• Standard Gaussian graphical models assume homogeneity, which is often violated in real-world data from diverse populations.
• Heterogeneity can lead to unobserved nonlinear relationships among variables, challenging traditional modeling approaches.
• Existing methods may fail to capture the nuanced conditional independence structures present in mixed populations.

Purpose of the Study:

• To develop and explore mixture models for Gaussian graphical models to address population heterogeneity.
• To enable the identification of distinct conditional independence structures within homogeneous subgroups.
• To provide a framework for analyzing complex, nonlinear relationships in heterogeneous data.

Main Methods:

• Exploration of infinite mixtures of Gaussian graphical models.
• Application of infinite hidden Markov models with Gaussian graphical model emission distributions.
• Clustering of heterogeneous populations into homogeneous groups, each with a unique graphical structure.

Main Results:

• The proposed mixture models effectively partition heterogeneous data into distinct clusters.
• Each identified cluster exhibits its own specific conditional independence structure.
• Analysis of pre-Euro foreign exchange rate data revealed significant trends and group-specific dynamics.

Conclusions:

• Mixture Gaussian graphical models offer a powerful approach for analyzing data from heterogeneous populations.
• These models accurately capture varying conditional independence structures across different subgroups.
• The methodology provides valuable insights into financial market dynamics, such as foreign exchange rate fluctuations.