A0353
Title: Minorization-maximization-based estimation for network models with parameter vectors of increasing dimension
Authors: Cornelius Fritz - Trinity College Dublin (Ireland) [presenting]
Michael Schweinberger - Pennsylvania State University (United States)
David Hunter - Pennsylvania State University (United States)
Abstract: Large and complex network data necessitate complex models accommodating local and global dependence. Local dependence refers, e.g., to transitive clustering based on common partners, implying the knowledge about other population members' connections. On the other hand, global dependence governs the general propensity to interact with other population members regardless of sharing a common neighborhood. A general framework is introduced to capture both types of dependencies that give rise to high-dimensional models for network data. Standard algorithms to tackle such computational problems are based on Metropolis-Hastings Monte Carlo or Newton-Raphson Methods, which perform poorly in high-dimensional settings. A minorization-maximization (MM) method is introduced for convex objective functions to alleviate this scaling issue. Quasi-Newton acceleration methods are employed to speed up the convergence of the algorithm. Moreover, a penalty is introduced to bypass issues of unidentifiable coefficients. In several applications, the performance of the algorithms is exhibited in comparison to currently available algorithms.
