Learning Exponential Random Graph Models

摘要: Exponential Random Graphs are common, simple statistical models for social network and other structures. Unfortunately, inference learning with them is hard networks larger than 20 nodes because their partition functions intractable to compute precisely. In this paper, we introduce a novel linear-time deterministic approximation these functions. Our main insight enabling advance that subgraph statistics sufficient derive lower bound The proposed method differs from existing methods in the way it exploits asymptotic properties of statistics. comparison current Monte Carlo simulation based methods, new scalable, stable, precise enough tasks. We show strengths approach experimentally theoretically.