jeek: A Fast and Scalable Joint Estimator for Integrating Additional Knowledge in Learning Multiple Related Sparse Gaussian Graphical Models

Beilun Wang, Yanjun Qi, Zhaoyang WangOpenalex Unknown 2018 · DOI · OpenAlex

Abstract

The jeek package provides a fast and scalable joint estimator (JEEK) for integrating additional knowledge into the learning of multiple related sparse Gaussian Graphical Models (sGGMs). The JEEK algorithm is designed to estimate multiple related precision matrices at scale, making it ideal for identifying complex networks—such as gene interactions—from high-dimensional, multi-context datasets. By performing data-driven network inference from heterogeneous data, this tool helps researchers translate aggregated data into meaningful graphs among entities.

The Challenge of Large-Scale Network Inference

In fields like bioinformatics and neuroscience, researchers often need to infer networks of relationships (graphs) from high-dimensional data. For example, understanding how thousands of genes interact across different tissues requires estimating multiple related networks. While these networks can be estimated individually, joint estimation is preferred because related tasks (e.g., different cancer stages or tissue types) likely share a common underlying structure.

However, existing joint estimators face two major hurdles:

  1. Scalability: As the number of features ($p$) and tasks ($K$) grows, traditional likelihood-based methods become computationally prohibitive, often scaling at $O(p^5)$.
  2. Prior Knowledge Integration: Most tools struggle to incorporate existing domain knowledge (such as known protein-protein interactions) into the statistical estimation process.

The JEEK Methodology

The core of the package is the Joint Elementary Estimator incorporating additional Knowledge (JEEK). Unlike traditional methods, JEEK uses a closed-form, entry-wise parallelizable solution that reduces computational complexity from $O(p^5)$ to $O(p^2)$.

Key technical innovations include:

  • Hybrid Norm Objective: The algorithm minimizes a novel hybrid norm that enforces a "superposition" of two weighted sparsity constraints—one for interactions shared across all tasks and another for task-specific patterns.
  • Knowledge Integration: Users can provide weight matrices ($W$) where specific values signal the likelihood of a connection based on prior knowledge. This acts as a "soft constraint," allowing the data to override the prior if evidence is strong enough.
  • Nonparanormal Support: Beyond standard Gaussian models, jeek can estimate "nonparanormal" models using Kendall’s tau correlation, providing robustness against non-Gaussian data distributions.
Parameters for the jeek function, including the weight matrix W for prior knowledge and covType for model selection.

Datasets and Benchmarking

The package includes several datasets for demonstration and benchmarking:

  • cancer: A microarray dataset of 133 breast cancer patients across 26 genes, split by clinical response.
  • nip_37_data: A preprocessed version of the NIPS Conference Papers dataset (1987–2015), featuring word counts for 37 specific terms across two time-based tasks.
  • exampleData: Simulated toy datasets used to demonstrate the package's ability to recover known "ground truth" shared graphs.

Key Insights and Results

  • Computational Breakthrough: By moving to an entry-wise approach, JEEK improves efficiency from $O(p^5 K^4)$ to $O(p^2 K^4)$. This allows the tool to handle "large $p$, large $K$" scenarios that were previously untouchable.
  • Statistical Efficiency: Despite its speed, JEEK achieves the same convergence rate—$O(\log(Kp)/n_{tot}$)—as much slower, state-of-the-art estimators.
  • Ease of Use: The R package includes built-in visualization tools and functions to export results directly into igraph objects for further network analysis.
Visualization parameters for the plot.jeek function, allowing for detailed graph analysis.

Limitations and Future Directions

Like most sGGM estimators, JEEK assumes underlying networks are relatively sparse; performance may degrade if true networks are very dense. Additionally, while the package allows for the integration of domain knowledge via weights ($W$), users must currently define these weights manually based on external expertise. Future iterations may explore automated weight tuning or GPU acceleration to further leverage the algorithm's parallel nature.

For further details, please refer to the original paper: Beilun Wang, Arshdeep Sekhon, Yanjun Qi "A Fast and Scalable Joint Estimator for Integrating Additional Knowledge in Learning Multiple Related Sparse Gaussian Graphical Models" (ICML 2018).

Summary updated

Abstract

Provides a fast and scalable joint estimator for integrating additional knowledge in learning multiple related sparse Gaussian Graphical Models (JEEK). The JEEK algorithm can be used to fast estimate multiple related precision matrices in a large-scale. For instance, it can identify multiple gene networks from multi-context gene expression datasets. By performing data-driven network inference from high-dimensional and heterogeneous data sets, this tool can help users effectively translate aggregated data into knowledge that take the form of graphs among entities. Please run demo(jeek) to learn the basic functions provided by this package. For further details, please read the original paper: Beilun Wang, Arshdeep Sekhon, Yanjun Qi "A Fast and Scalable Joint Estimator for Integrating Additional Knowledge in Learning Multiple Related Sparse Gaussian Graphical Models" (ICML 2018) &lt;<a href="https://doi.org/10.48550/arXiv.1806.00548" target="_top">doi:10.48550/arXiv.1806.00548</a>&gt;.
BibTeX
@inproceedings{wang2018jeek,
  author = {Beilun Wang and Yanjun Qi and Zhaoyang Wang},
  title = {{jeek: A Fast and Scalable Joint Estimator for Integrating Additional Knowledge in Learning Multiple Related Sparse Gaussian Graphical Models}},
  booktitle = {Openalex Unknown},
  year = {2018},
  doi = {10.32614/cran.package.jeek},
}
Figure