Upcoming: Marloes Maathuis, Ph.D.
Professor, Department of Statistics, ETH Zurich, Switzerland, “High-Dimensional Consistency in Score-Based and Hybrid Structure Learning ,” at 9:00 am on Friday, November 11, 2016, in Rooms 407A/B BAUM, 5607 Baum Blvd., The Offices at Baum.
Abstract: The main approaches for learning Bayesian networks can be classified as constraint-based, score-based or hybrid methods. Although high-dimensional consistency results are available for the constraint-based PC algorithm, such results have been lacking for score-based and hybrid methods, and most hybrid methods are not even proved to be consistent in the classical setting where the number of variables remains fixed. We study the score-based Greedy Equivalence Search (GES) algorithm, as well as hybrid algorithms that are based on GES. We show that such hybrid algorithms can be made consistent in the classical setting by using an adaptive restriction on the search space. This leads to the adaptively restricted GES (ARGES) algorithm. Moreover, we prove consistency of GES and ARGES for certain sparse high-dimensional scenarios. This is joint work with Preetam Nandy and Alain Hauser
Biography: Dr. Maathuis’ research areas include causal inference, graphical models, high-dimensional statistics, asymptotics, and application of statistics. She is currently associate editor for the Annals of Statistics, Biometrika, and the Scandinavian Journal of Statistics.
Lecturer, Department of Statistical Science and Centre for Computational Statistics and Machine Learning, University College London, “Learning Causal Effects: Bridging Instruments and Backdoors,” at 11:00 am on Thursday, September 15, 2016, in Rooms 407A/B BAUM, 5607 Baum Blvd., The Offices at Baum.
Abstract: We consider the problem of learning the causal effect of some treatment X on some outcome Y knowing that there is a background set of variables that are not caused by either. We first discuss what can be done in linear models, when unmeasured confounding between X and Y cannot be blocked and candidate instrumental variables are proposed from testable constraints in the observed distributions. A characterization of what can be discovered is given, including limitations, equivalence classes and to which extent non-Gaussianity assumptions can help. In the second half, we generalize algorithms that find backdoor adjustment sets exploiting the faithfulness assumption. The idea is to provide a whole continuum of relaxations of faithfulness, from which we will show how algorithms for learning backdoor adjustments can provide instrumental variables that give bounds on causal effects for discrete distributions.
Joint work with Shohei Shimizu and Robin Evans.
Biography: Dr. Ricardo Silva is a Senior Lecturer in the Department of Statistical Science, Adjunct Faculty in the Gatsby Computational Neuroscience Unit, and in the management group of the Centre for Computational Statistics and Machine Learning (CSML) at UCL. He is also in the management group of the EPSRC Network on Computational Statistics and Machine Learning and a fellow of the Alan Turing Institute. Dr Silva has an extensive experience in research in machine learning, in particular in areas such as graphical models, latent variable models and causality. During his PhD at Carnegie Mellon University, Dr Silva laid off some early work on causal structure identification for models with unobserved variables. Dr Silva introduced new approaches for graphical model construction and inference, models for network data in prediction problems and information retrieval and developed inference algorithms for complex distributions and causal models, among other contributions.
Jun Zhu, PhD, Professor, Icahn Institute and Department of Genetics and Genomic Sciences, Icahn School of Medicine at Mount Sinai