Hub Detection in Gaussian Graphical Models

TL;DR

This paper introduces a new method called IPC-HD for directly detecting hubs in Gaussian graphical models by leveraging spectral decomposition, improving accuracy and speed over existing methods, and demonstrating its effectiveness on gene expression data.

Contribution

The paper proposes a novel spectral-based approach for hub detection in GGMs that bypasses graph estimation, with proven consistency and superior performance.

Findings

IPC-HD outperforms existing methods in simulations.

The method is computationally efficient.

Application identifies biologically relevant hub genes.

Abstract

Graphical models are popular tools for exploring relationships among a set of variables. The Gaussian graphical model (GGM) is an important class of graphical models, where the conditional dependence among variables is represented by nodes and edges in a graph. In many real applications, we are interested in detecting hubs in graphical models, which refer to nodes with a significant higher degree of connectivity compared to non-hub nodes. A typical strategy for hub detection consists of estimating the graphical model, and then using the estimated graph to identify hubs. Despite its simplicity, the success of this strategy relies on the accuracy of the estimated graph. In this paper, we directly target on the estimation of hubs, without the need of estimating the graph. We establish a novel connection between the presence of hubs in a graphical model, and the spectral decomposition of the underlying covariance matrix. Based on this connection, we propose the method of inverse principal components for hub detection (IPC-HD). Both consistency and convergence rates are established for IPC-HD. Our simulation study demonstrates the superior performance and fast computation of the proposed method compared to existing methods in the literature in terms of hub detection. Our application to a prostate cancer gene expression dataset detects several hub genes with close connections to tumor development.