BISS workshop, June 12, 2005

Building causal models from observational data — an overview from an artificial intelligence perspective

Eric Neufeld (University of Saskatchewan) Sunday, 12 June 2005 — 13:00-17:15

From its beginning, the Artificial Intelligence (AI) community has had an interest in theories of causality, despite the philosophical difficulties associated with it. Intelligent robots and software must be able to effectively infer, and reason with, practical knowledge about cause and effect. This holds whether this agent is physically exploring an environment or just providing advice.

The results of Judea Pearl at UCLA and Spirtes, Glymour and Scheines (SGS) at CMU in the 1990s put causal inference on statistical and probabilistic foundations. These groups independently discovered algorithms for inferring causal structure from observational data, and developed methodologies for applying these techniques to a variety of real world problems. Since then, this area has matured. It has gained some respect from the mainstream statistics community, but it has also drawn criticisms.

The purpose of this workshop is to introduce participants to the philosophical and statistical underpinnings of causality, beginning with basic definitions and then using these definitions iteratively to construct causal structure from raw data. The basic algorithm identifies potential causal relations and ‘genuine’ causal relations, and can also identify spurious relationships. These structures can be used to compute the results of interventions (actions) in domains, including causal effects of a variable in the presence of hidden/unmeasurable confounders. Causal structures also provide a possible ‘solution’ to Simpson’s paradox. We review some of the claimed successes of this approach, and also discuss critical response (and response to the criticisms) of this work.