Gaussian Graphical Modelling Without Independence Assumptions for Uncentered Data

Bailey Andrew, David R. Westhead, Luisa CutilloAAAI 2025 · DOI

TL;DR

This paper tackles a critical limitation in Gaussian graphical modeling: the assumption of zero-mean data, which can lead to drastically incorrect inferences. The authors introduce a novel "Kronecker-sum-structured mean" assumption and a corresponding optimization algorithm that effectively estimates non-zero means, significantly improving model performance on various datasets, including gene expression data.

What Problem Does This Paper Address?

Gaussian graphical models are powerful tools for uncovering relationships between variables, often represented as networks or graphs. A key assumption in many of these models, particularly those dealing with multi-axis or tensor-valued data (like gene expression data where you have cells and genes), is that the data has zero mean. This assumption simplifies computations but can be problematic. When data is not truly zero-mean, forcing it to be can introduce significant errors, essentially distorting the underlying relationships the model is trying to find. This is especially true for complex, multi-dimensional datasets where the mean itself might carry important structural information. The paper aims to relax this restrictive zero-mean assumption to build more robust and accurate graphical models.

Methods

The core of the paper's contribution lies in introducing and addressing the "Kronecker-sum-structured mean" assumption. Instead of assuming a zero mean, they propose a mean that can be decomposed in a similar structured way as the precision matrix (which represents the graph's structure). This decomposition allows for the estimation of a non-zero mean that is compatible with the Kronecker-sum structure of the precision matrix.

To estimate the parameters of this noncentral model (i.e., a model with a non-zero mean), they develop an optimization strategy. While the joint optimization problem of estimating both the mean and the precision matrix is non-convex, they show it's "biconvex," meaning it's convex with respect to each parameter individually when the other is fixed. This allows them to use a block coordinate descent approach: they iteratively optimize the mean parameters and the precision matrix parameters. They also demonstrate that despite the non-convexity, the problem has a unique global optimum, and their algorithm converges to it.

Datasets

The paper evaluates their method on several synthetic datasets generated from Kronecker-sum normal distributions with various mean structures (constant offset, independent Gaussian/Poisson noise, structured means). They also test on real-world datasets:

  • COIL-20 video dataset: Used to demonstrate improvements in object recognition tasks by modeling connections between video frames.
  • E-MTAB-2805 scRNA-seq dataset: A single-cell RNA sequencing dataset used to show the impact of mean correction on gene regulatory network inference, particularly in identifying cell cycle stage relationships.

Key Insights and Ideas

  • The zero-mean assumption is a significant bottleneck: The paper convincingly shows that ignoring the mean can lead to inferences that are not just inaccurate but actively misleading, even identifying repulsive forces where attractive ones should exist.
  • Structured means are crucial: Just as structured precision matrices capture dependencies, structured means can capture underlying systematic shifts or biases in the data, which are important for accurate modeling.
  • Biconvexity enables optimization: The problem's biconvex nature is key to developing a practical optimization algorithm, allowing for iterative refinement of mean and precision estimates.
  • Unique global optimum guarantees reliability: The theoretical proof that a unique global optimum exists for the noncentral model is a strong theoretical underpinning, assuring users that the algorithm will find the best possible solution.

Key Results

  • On synthetic data, their noncentral approach significantly outperforms zero-mean methods, especially when data is generated with a non-zero mean offset. For example, in the COIL-20 dataset, their method improved the percentage of "correct" edges from around 40% to over 85%.
  • For the scRNA-seq dataset, their method corrected a strong, counter-intuitive repulsive force identified by zero-mean methods, revealing a more reasonable positive tendency for cells in the same cycle stage to connect.
  • The algorithm is presented as a "drop-in" wrapper, meaning it can be used with existing precision matrix estimation algorithms (like GmGM and TeraLasso) with minimal modification.

Limitations and Future Work

  • Identifiability: The paper notes that the initial parameterization of the mean and precision matrices might not be identifiable and proposes specific constraints (e.g., sum of means to zero) to ensure identifiability.
  • Computational Complexity: While the algorithm is efficient, the paper mentions that for very large datasets, further optimizations might be needed. The complexity is noted as O(I(G + M)), where G is graph estimation cost and M is mean estimation cost.
  • Future Work: The authors suggest that with minor modifications, their method can preserve the "only-one-eigendecomposition" property of certain algorithms like GmGM, further enhancing computational efficiency. They also hint at extending this work to non-Gaussian distributions.

Summary updated

BibTeX
@inproceedings{andrew2025gaussian,
  author = {Bailey Andrew and David R. Westhead and Luisa Cutillo},
  title = {{Gaussian Graphical Modelling Without Independence Assumptions for Uncentered Data}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i15.33689},
}
Figure