Exploratory Model Analysis: Interactive Graphical Methods for Model Selection and Comparison

• Activity Number: 86
• Type: Invited
• Date/Time: Monday, July 30, 2007 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Graphics
• Author(s): Simon Urbanek*+
• Companies: AT&T Labs - Research
• Keywords: exploratory model analysis ; model selection ; visualization ; tre emethods ; model ensembles

Abstract

Increasing availability of computing power widens the choice of models and methods in many fields of data analysis and data mining. However, the increasing variety also makes it harder to decide which methods to use for a given problem. Many publications have been written on model selection and comparison, mainly with the goal to select "the" optimal model, usually condensing the properties into a single number. In most cases this is not adequate as models have different properties with respect to the data that cannot be measured only from one viewpoint. In this talk we discuss a different approach that attempts to understand differences in statistical models by using visualization of both data and model information. We will show a general framework for exploratory model analysis and illustrate its use on practical examples with models ranging from linear models to tree methods.