Title: Model selection and local geometry
Authors: Robin Evans - University of Oxford (United Kingdom) [presenting]
Abstract: Model selection is a task of fundamental importance in statistics, and advances in high-dimensional model selection have been one of the major areas of progress over the past 20 years. Examples include covariate selection in linear regression, and models based on patterns of zeros in the inverse covariance matrix. Much of this progress has been due to penalized methods such as the lasso, and efficient methods for solving the relevant convex optimization problems. However in other classes, such as directed graphical models, correct model selection is provably hard. We give a geometric explanation for why standard convex penalized methods cannot be adapted to directed graphs, based on the local geometry of the different models at points of intersection. These results also show that it is `statistically' hard to learn these models, and that much larger samples will typically be needed for moderate effect sizes. This has implications for other types of graphical model selection, and especially for causal models, as well as time series models. We provide some relevant heuristics that give insights into the feasibility of model selection in various classes of graphical model, including ancestral graph models, LWF chain graph models, and nested Markov models.