Bayesian Inference for Latent Chain Graphs

Lu Deng, Maria De Iorio, Ajay Jasra, Gary L. Rosner

Abstract

In this article we consider Bayesian inference for partially observed Andersson-Madigan-Perlman (AMP) Gaussian chain graph (CG) models. Such models are of particular interest in applications such as biological networks and financial time series. The model itself features a variety of constraints which make both prior modeling and computational inference challenging. We develop a framework for the aforementioned challenges, using a sequential Monte Carlo (SMC) method for statistical inference. Our approach is illustrated on both simulated data as well as real case studies from university graduation rates and a pharmacokinetics study.

Introduction

Probabilistic graphical models typically rely on Directed Acyclic Graphs (DAGs) for asymmetric dependencies or Undirected Graphs for symmetric ones. However, many real-world systems—such as gene regulatory networks or financial markets—exhibit a mixture of both. Chain Graphs (CGs) provide a unifying framework for these dependencies. This paper addresses the challenge of latent structure inference: discovering the underlying graph when it is unobserved. Specifically, it focuses on the Andersson-Madigan-Perlman (AMP) interpretation, which requires satisfying strict topological constraints, such as the absence of semi-directed cycles.

Methodology

The authors propose a Bayesian framework for Gaussian chain graphs where the precision matrix is constrained by the graph's topology. The model utilizes a matrix $B$ for regression coefficients (directed edges) and a matrix $\Omega$ for residual precision (undirected edges).

To navigate the complex, discrete space of valid graphs, the researchers employ a Sequential Monte Carlo (SMC) sampler. Unlike standard MCMC, which can struggle with local optima in high-dimensional graph spaces, SMC uses a population of "particles" (candidate graphs) and gradually tempers the distribution from the prior to the posterior. This approach is more robust and naturally parallelizable.

SMC diagnostics showing Effective Sample Size (ESS), parameter traces, acceptance rates, and log-target density.

Experimental Results

The framework was evaluated on simulated data and two real-world case studies.

Simulated Data and Structure Recovery

In a benchmark test with 10 independent nodes, the algorithm successfully recovered the empty graph structure. Posterior probabilities for the "no edge" condition were estimated above 0.9 for most node pairs, demonstrating that the model does not falsely identify connections where none exist.

University Graduation Rates

The model was applied to a dataset of 8 variables (e.g., faculty salary, student-teacher ratio, graduation rate) from 159 US universities. The resulting "Empirical Graph" captured complex dependencies that outperformed traditional Structural Equation Models (SEM). The chain graph achieved superior model selection metrics, with an AIC of 61.37 and a BIC of -50.52.

The empirical graph and its corresponding adjacency matrix for the university graduation rates dataset.

Pharmacokinetics Study

In a study of Tenofovir concentrations ($n=11, p=7$), the model recovered known biological associations between blood plasma and rectal tissue levels. This success in a data-starved regime highlights the power of the Bayesian prior in identifying meaningful structures from small sample sizes.

Conclusion

This Bayesian framework for latent chain graphs offers a flexible alternative to DAGs, particularly when the direction of causality is not strictly asymmetric. While the SMC sampler provides a powerful tool for navigating constrained graph spaces, future work will focus on improving scalability for high-dimensional data and exploring computational approximations to handle larger networks more efficiently.

Summary updated

Abstract

In this article we consider Bayesian inference for partially observed Andersson-Madigan-Perlman (AMP) Gaussian chain graph (CG) models. Such models are of particular interest in applications such as biological networks and financial time series. The model itself features a variety of constraints which make both prior modeling and computational inference challenging. We develop a framework for the aforementioned challenges, using a sequential Monte Carlo (SMC) method for statistical inference. Our approach is illustrated on both simulated data as well as real case studies from university graduation rates and a pharmacokinetics study.
BibTeX
@inproceedings{deng2020bayesian,
  author = {Lu Deng and Maria De Iorio and Ajay Jasra and Gary L. Rosner},
  title = {{Bayesian Inference for Latent Chain Graphs}},
  booktitle = {Foundations},
  year = {2020},
  doi = {10.3934/fods.2020003},
}
Figure