[Paper Review] An Annotated Graph Model with Differential Degree Heterogeneity for Directed Networks

Stefan Stein, Chenlei Leng|arXiv (Cornell University)|Aug 21, 2021

Complex Network Analysis Techniques59 references4 citations

TL;DR

This paper proposes a novel annotated graph model for directed networks that explicitly models differential degree heterogeneity and covariate effects using a sparse, penalized likelihood approach with ℓ₁-regularization. The method ensures estimation and selection consistency under sparsity, enabling valid inference on covariate effects without debiasing, and is validated through theory and data on real-world networks like the lawyer friendship network.

ABSTRACT

Directed networks are conveniently represented as graphs in which ordered edges encode interactions between vertices. Despite their wide availability, there is a shortage of statistical models amenable for inference, specially when contextual information and degree heterogeneity are present. This paper presents an annotated graph model with parameters explicitly accounting for these features. To overcome the curse of dimensionality due to modelling degree heterogeneity, we introduce a sparsity assumption and propose a penalized likelihood approach with $\ell_1$-regularization for parameter estimation. We study the estimation and selection consistency of this approach under a sparse network assumption, and show that inference on the covariate parameter is straightforward, thus bypassing the need for the kind of debiasing commonly employed in $\ell_1$-penalized likelihood estimation. Simulation and data analysis corroborate our theoretical findings.

Motivation & Objective

• Address the lack of statistical models for directed networks that incorporate both contextual covariates and degree heterogeneity.
• Overcome the curse of dimensionality from modeling individual node-specific degree heterogeneity parameters in large networks.
• Develop a method that enables consistent estimation and variable selection for covariate effects in sparse directed networks.
• Ensure valid statistical inference on covariate parameters without requiring debiasing procedures common in ℓ₁-regularized models.
• Demonstrate the model's theoretical and empirical performance on real-world network data, such as the Lazega lawyer friendship network.

Proposed method

• Proposes a directed graph model where edge probabilities depend on node-specific out-degree and in-degree heterogeneity parameters (αᵢ, βⱼ), a global density parameter (μ), and covariates (Zᵢⱼ) through a logistic link function.
• Imposes sparsity on the degree heterogeneity parameters α and β to reduce model dimensionality and enable estimation in large networks.
• Applies ℓ₁-regularization to the likelihood function to achieve simultaneous estimation and variable selection, with identifiability maintained by setting min(αᵢ) = min(βⱼ) = 0.
• Establishes estimation and selection consistency under a sparse network assumption, proving that the true model structure is recovered with high probability.
• Demonstrates that inference on the covariate parameter γ is straightforward due to the absence of bias in the regularized estimator, bypassing the need for debiasing steps common in high-dimensional models.
• Validates the model using simulation studies and analysis of the Lazega lawyer friendship network, showing strong empirical performance and consistency with theoretical results.

Experimental results

Research questions

• RQ1 Can a statistical model for directed networks jointly account for degree heterogeneity and covariate effects while remaining computationally feasible in high-dimensional settings?
• RQ2 Does ℓ₁-regularization in a penalized likelihood framework yield consistent estimation and variable selection for both degree heterogeneity and covariate parameters in sparse directed networks?
• RQ3 Can inference on the covariate effect parameter γ be performed without debiasing procedures, given the structure of the model and regularization?
• RQ4 To what extent does the model capture real-world network features such as homophily and power-law degree distributions?
• RQ5 How well does the model perform on real data, such as the Lazega lawyer friendship network, in terms of both fit and interpretability?

Key findings

• The proposed model achieves estimation and selection consistency for both degree heterogeneity and covariate parameters under a sparse network assumption.
• The ℓ₁-regularized likelihood estimator for the covariate parameter γ is asymptotically normal and does not require debiasing, enabling straightforward inference.
• The model can generate power-law degree distributions under appropriate parameter configurations, consistent with real-world network features.
• The simulation studies confirm that the method correctly identifies relevant covariates and sparsity patterns in the degree heterogeneity parameters.
• Empirical analysis of the Lazega lawyer network shows that covariates such as age, gender, and office location significantly influence tie formation, with homophily effects detectable through positive γ estimates.
• Theoretical results show that the empirical degree distribution converges in probability to a power law with exponent τ = 2 under the specified conditions, supporting the model's ability to reproduce real network structures.