[Paper Review] ergm 4: Computational Improvements

Pavel N. Krivitsky, David R. Hunter+2 more|arXiv (Cornell University)|Mar 15, 2022

Complex Network Analysis Techniques4 citations

TL;DR

This paper presents major computational improvements in the ergm 4.0 package for fitting exponential-family random graph models (ERGMs) in R, focusing on optimized Markov chain Monte Carlo (MCMC) and simulated annealing (SAN) algorithms. The enhancements deliver up to several orders of magnitude speedup, enabling feasible analysis of larger and more complex networks than previously possible with earlier versions.

ABSTRACT

The ergm package supports the statistical analysis and simulation of network data. It anchors the statnet suite of packages for network analysis in R introduced in a special issue in Journal of Statistical Software in 2008. This article provides an overview of the performance improvements in the 2021 release of ergm version 4. These include performance enhancements to the Markov chain Monte Carlo and maximum likelihood estimation algorithms as well as broader and faster searching for networks with certain target statistics using simulated annealing.

Motivation & Objective

• To address the computational inefficiency of ERGM estimation in large or complex networks, particularly due to intractable normalizing constants and slow MCMC convergence.
• To improve the performance of Markov chain Monte Carlo (MCMC) algorithms used for network simulation and likelihood estimation in ERGMs.
• To accelerate the simulated annealing (SAN) process for finding networks matching target statistics, especially under constraints.
• To enable efficient maximum likelihood estimation (MCMLE) with bounded degree constraints and large network sizes.
• To support scalable, high-performance network modeling in the statnet suite by integrating algorithmic and implementation-level optimizations.

Proposed method

• Optimized Metropolis–Hastings MCMC algorithms with improved proposal mechanisms and reduced computational overhead per iteration.
• Introduced adaptive MCMC interval selection to balance mixing and computational cost, improving convergence efficiency.
• Enhanced simulated annealing (SAN) with faster convergence to target statistics by improving search heuristics and parallelization.
• Implemented bounded degree constraints via the `bd()` function, allowing efficient sampling of networks with specified vertex degree limits.
• Integrated parallelization across MCMC chains and SAN steps to leverage multi-core processors and reduce wall-clock time.
• Optimized internal data structures and memory access patterns in C++ to reduce overhead in network manipulation and sufficient statistic computation.

Experimental results

Research questions

• RQ1 How can MCMC algorithms in ERGM estimation be made significantly faster without compromising accuracy?
• RQ2 To what extent do adaptive MCMC intervals improve convergence and reduce computation time in large network models?
• RQ3 Can simulated annealing be accelerated to efficiently find networks matching complex target statistics, especially under structural constraints?
• RQ4 How do bounded degree constraints affect computational performance, and can they be efficiently enforced in large-scale ERGM fitting?
• RQ5 What performance gains are achievable in modern hardware through algorithmic and implementation-level optimizations in the ergm 4.0 package?

Key findings

• The MCMC and SAN algorithms in ergm 4.0 achieve up to several orders of magnitude improvement in computational speed compared to version 3.10.
• Model fitting with bounded degree constraints that previously took over 5 hours in version 3.10 now completes in under 10 minutes in ergm 4.1, representing a 30x speedup.
• Using a 10,000-node network, ergm 4.1 with adaptive MCMC intervals completed in 569 seconds, compared to 22.64 hours (81,500 seconds) in version 3.10 with short intervals.
• The use of parallelized MCMC and SAN significantly reduces wall-clock time, with performance scaling effectively across multiple CPU cores.
• The integration of efficient C++ code and optimized data structures reduced memory access overhead, contributing to faster network generation and statistic evaluation.
• The new implementation enables the analysis of network sizes and model complexities that were previously infeasible due to computational constraints.