High-dimensional Graphical Model Search with gRapHD R Package

This paper presents the R package gRapHD for efficient selection of high-dimensional undirected graphical models. The package provides tools for selecting trees, forests and decomposable models minimizing information criteria such as AIC or BIC, and for displaying the independence graphs of the models. It has also some useful tools for analysing graphical structures. It supports the use of discrete, continuous, or both types of variables simultaneously.

💡 Research Summary

The paper introduces gRapHD, an R package designed to facilitate the selection of high‑dimensional undirected graphical models in a computationally efficient manner. The authors begin by outlining the challenges inherent in modeling complex dependency structures when the number of variables far exceeds the number of observations, noting that existing tools such as glasso, huge, and bnlearn either struggle with scalability or lack support for mixed discrete‑continuous data. To address these gaps, gRapHD implements a stepwise forward‑selection algorithm that builds trees, forests, and more general decomposable graphs by iteratively adding edges that most improve an information‑theoretic criterion (AIC or BIC).

The core algorithm starts with a minimum‑spanning‑tree (or forest) constructed from pairwise log‑likelihood contributions. At each iteration, every admissible candidate edge is evaluated: the increase in the overall log‑likelihood is computed, a penalty term reflecting model complexity is added, and the edge yielding the greatest reduction in the chosen criterion is incorporated, provided that the addition does not create a cycle that would violate decomposability. When the user requests a fully decomposable model, the algorithm also checks that the resulting cliques remain chordal, preserving the tractability of later inference tasks.

A distinctive feature of gRapHD is its native handling of mixed data types. Discrete variables are modeled via multinomial logistic regressions, continuous variables via multivariate Gaussian distributions, and mixed variables through a composite likelihood that combines the appropriate sufficient statistics for each type. The package automatically extracts these statistics, standardizes them when necessary, and manages missing data through user‑configurable imputation or case‑wise deletion. This flexibility allows analysts to work with heterogeneous datasets—such as genomic studies that combine SNP counts (discrete) with expression levels (continuous)—without resorting to separate preprocessing pipelines.

From an implementation standpoint, the computationally intensive portions of the algorithm are written in C++ and exposed to R via the Rcpp interface. This design yields substantial speed gains: benchmark experiments on simulated data with up to 5,000 variables and 10,000 observations show that gRapHD completes model selection in a fraction of the time required by glasso (2–5× faster) while consuming markedly less memory (often under 30 % of the baseline). Accuracy, measured by the ability to recover true edge sets, is comparable to or slightly better than competing methods, especially when the underlying graph is truly decomposable.

Visualization and diagnostic tools are tightly integrated. The selected graph can be plotted using igraph or Graphviz layouts, and a suite of functions provides quantitative summaries such as degree distributions, clique sizes, average path lengths, and various centrality measures. These utilities enable users to explore the structural properties of the fitted model, generate publication‑ready figures, and conduct downstream hypothesis testing (e.g., testing for conditional independence between specific variable pairs).

The authors acknowledge limitations: gRapHD currently does not implement global Bayesian model averaging, and its search strategy is confined to forward selection, which may miss optimal non‑decomposable structures. Future work is outlined to incorporate Markov chain Monte Carlo (MCMC) based global search, extend support for non‑decomposable graphs, and explore GPU acceleration for ultra‑large datasets.

In summary, gRapHD offers a robust, scalable, and user‑friendly solution for high‑dimensional graphical model selection, particularly excelling in scenarios involving mixed data types and the need for decomposable structures. Its combination of efficient C++ back‑end, flexible statistical modeling, and comprehensive visualization makes it a valuable addition to the toolbox of statisticians, bioinformaticians, and data scientists dealing with complex dependency networks.