[Paper Review] Direct Learning of Sparse Changes in Markov Networks by Density Ratio Estimation

Song Liu, John A. Quinn+3 more|arXiv (Cornell University)|Apr 25, 2013

Statistical Methods and Inference18 references4 citations

TL;DR

This paper proposes a direct method for learning sparse structural changes in Markov networks by estimating the density ratio between two data distributions, avoiding computationally expensive separate model fitting. By formulating the problem as density ratio estimation with sparsity-inducing regularization, the method efficiently identifies structural changes while circumventing normalization constant computation, and achieves improved scalability through dual optimization.

ABSTRACT

We propose a new method for detecting changes in Markov network structure between two sets of samples. Instead of naively fitting two Markov network models separately to the two data sets and figuring out their difference, we \emph{directly} learn the network structure change by estimating the ratio of Markov network models. This density-ratio formulation naturally allows us to introduce sparsity in the network structure change, which highly contributes to enhancing interpretability. Furthermore, computation of the normalization term, which is a critical bottleneck of the naive approach, can be remarkably mitigated. We also give the dual formulation of the optimization problem, which further reduces the computation cost for large-scale Markov networks. Through experiments, we demonstrate the usefulness of our method.

Motivation & Objective

• To address the limitation of indirect structural change detection in Markov networks, where separate modeling of two networks leads to computational and interpretability issues.
• To enable direct learning of structural changes without fitting two full Markov network models, reducing computational cost and improving interpretability.
• To incorporate sparsity in the change detection process, favoring only meaningful, sparse structural differences between data distributions.
• To eliminate the need for normalization constant computation, a major bottleneck in maximum likelihood estimation of Markov networks.
• To develop a scalable optimization framework via dual formulation, enabling application to large-scale Markov networks.

Proposed method

• Formulates structural change detection as a density ratio estimation problem between two data distributions, bypassing separate Markov network fitting.
• Uses a pairwise Markov network model with a log-linear exponential family form to represent the density ratio.
• Applies ℓ1 regularization on the difference of network parameters to enforce sparsity in the structural changes.
• Derives a dual optimization problem that decouples the computation and reduces the dimensionality of the optimization task.
• Employs a Lagrangian relaxation and sub-gradient method to solve the non-differentiable ℓ1-regularized dual problem.
• Utilizes the fact that the normalization constant in the log-partition function can be approximated via importance sampling or analytical forms, avoiding direct computation.

Experimental results

Research questions

• RQ1 Can structural changes in Markov networks be learned directly without separately fitting two full models?
• RQ2 Can density ratio estimation effectively capture sparse structural changes while avoiding normalization constant computation?
• RQ3 Does the proposed direct method outperform indirect two-step approaches in terms of accuracy, sparsity, and computational efficiency?
• RQ4 Can the dual formulation significantly reduce computational cost for large-scale Markov networks?
• RQ5 How does the method perform under non-Gaussian and nonparanormal distributions where traditional MLE-based methods fail?

Key findings

• The proposed method successfully learns sparse structural changes directly from data without fitting two separate Markov networks, improving interpretability.
• By avoiding explicit computation of normalization constants, the method achieves significant computational savings compared to maximum likelihood estimation.
• The dual formulation reduces the optimization dimensionality and enables efficient solution of large-scale problems.
• The method achieves comparable or better performance than existing fused-lasso approaches on synthetic and real-world data, especially when changes are sparse but individual networks are dense.
• Empirical results demonstrate that the method is robust to non-Gaussian and nonparanormal distributions, extending beyond the limitations of Gaussian-based methods.
• The use of density ratio estimation allows for consistent estimation of structural changes even when the underlying distributions are complex or high-dimensional.