Efficient Markov Chain Monte Carlo Estimation of Exponential Random Graph models Maksym Byshkin1, Alex Stivala2, Antonietta Mira1,3, Garry Robins2, Alessandro Lomi1,2 1Università della Svizzera italiana, Lugano, Switzerland 2University of Melbourne, Australia 3Università dell’Insubria, Como, Italy Third European Conference on Social Networks, Mainz, Germany, September 26 - 29, 2017 Introduction We have developed efficient Markov chain-based algorithm to perform Maximum Likelihood parameter estimation for probability distributions with intractable normalizing constants ERGMs are exponential family of probability distributions for dependent network data 1 æ ö å p(x,q) = exp q z (x) ç ÷ A A k è ø A æ ö å å k = exp q z (x) ç ÷ A A è ø x A z (x) are networks statistics (e.g. number of ties, triangles, stars ..) A Robins G, Snijders T, Wang P, Handcock M, & Pattison P (2007) Recent developments in exponential random graph (p*) models for social networks. Social Networks 29(2):192-215. Introduction We want to find MLE of the model parameters q ( ) E z (x) = z (x ) p(q) A A obs å are expected statistics E (z (x)) = z (x)p(x,q) p(q) A A x Metropolis-Hastings algorithm may be used to compute E (z (x)) and to generate simulated data x(q) p(q) A And we want to solve inverse problem: find q(x ) obs Lehmann, E.L., Casella, G. Theory of point estimation (2006) Existing estimation approaches 1) Bayesian 2) Geyer-Thompson MCMCMLE 3) Method of Moments (Stochastic approximation) - Iteratively update q A q - At many different values perform MCMC A simulations to draw simulated data x(q ) A MLE is computationally expensive We can find MCMCMLE without MCMC simulation Metropolis-Hastings algorithm Given the state x the state x’ is proposed with probability q(x ® x ') ì q(x' ® x)p(x',q )ü Acceptance probability a(x ®=x',q ) min 1, A í ý A q(x ® x')p(x,q ) î þ A Transition probability P(x ®®=x',q ) ®q(x x')a(x x',q ) A A The algorithm generates a Markov chain x that converges to 𝜋 𝑥, 𝜃 t if the algorithm step t is larger then the burn-in time Metropolis N, Rosenbluth AW, Rosenbluth MN, Teller AH, & Teller E Equation of state calculations by fast computing machines. The journal of chemical physics 21, 1087-1092 (1953) Hastings WK, Monte Carlo sampling methods using Markov chains and their applications. Biometrika 57, 97-109 (1970) New approaches for MLE 𝜋(𝑥, 𝜃) I can show that for simulated networks drawn from MLE may be found from Equilibrium Expectation (EE): ( ) E dz (x,q) = 0 p(q) A � - - 𝑑𝑧 (𝑥, 𝜃) = + P 𝑥 → 𝑥 , 𝜃 (𝑧 𝑥 −𝑧 𝑥 ) 𝐴 1 1 0 / The left part may be computed by Monte Carlo integrations, without time consuming MCMC simulation New approaches for MLE If observed network x is large then one sample is enough obs ˆ å ( ) P(x ® x ',q) z (x=') - z (x ) 0 obs A A obs x' But may be used also for small networks x , x ,.., x If we have data samples iid from 𝜋 ( 𝑥 , 𝜃 ) : S S S 1 2 n � 1 𝐸 (dz (𝑥, 𝜃)) = + dz (𝑥 , 𝜃) 4 5 𝐴 𝐴 : 𝑛 ; ; Very fast MLE ! MLE for empirical data How to apply this methodology for empirical data x ? obs We can draw simulated data x so that z (x) = z (x ) A A obs å ( ) P(x ® x',q ) z=(x') - z (x) 0 A A A ( ) E z (x) = z (x ) x' q A A obs z (x) = z (x ) A A obs We have developed MCMC algorithm to solve this system of equation EE algorithm for empirical data MCMC to constrain the values of all the statistics EE algorithm for empirical data MoM (Snijders 2002) and MCMCMLE (Geyer CJ & Thompson 1992) require many converged outputs of Metropolis-Hastings algorithm EE algorithm does not need such outputs. Instead it generates one converged output EE algorithm is similar to Metropolis-Hastings algorithm, but allows Monte Carlo simulation to be performed while constraining the values of statistics z (x) and in such a way that the EE condition is satisfied A
Description: