A0605
Title: Extending linear programming ecological inference methods by machine learning
Authors: Jose M Pavia - Universitat de Valencia (Spain) [presenting]
Abstract: Ecological inference (EI) is devised to forecast unknown inner-cells of two-way contingency tables by inferring conditional distribution probabilities. This outlines one of the more conspicuous and long-standing social science problems present in many areas, with political science and sociology being the disciplines where they are chiefly more frequent. For instance, EI algorithms are used to estimate vote transfer matrices between elections, infer split-ticket voting behaviors or reveal social and racial voting patterns. In the last years, we have experienced an explosion of methods to solve these problems from the Bayesian approach, based on a hierarchical multinomial-Dirichlet Bayesian model. The use of these methods, however, requires highly trained analysts and usually entails high computational costs. Recently, a new family of algorithms that considerably simplify the resolution of these problems has been proposed based on mathematical programming. The first wave of these new algorithms, which are at least as accurate as the Bayesian-based algorithms, are available in the R-package lphom. The goal is to show the accuracy improvements that we are achieving by integrating statistical learning approaches in this new methodology. Specifically, through the use of new algorithms based on (inspired by) bagging, genetic boosting and reinforcement learning.
