Speaker Judea Pearl and Elias Bareinboim
Host Joaquin Quiñonero Candela
Date recorded 5 December 2013
The traditional aim of machine learning methods is to infer meaningful features of an underlying probability distribution from samples drawn of that distribution. With the help of such features, one can infer associations of interest and predict or classify yet unobserved samples. Causal analysis goes one step further; it aims at inferring features of the data-generating process, that is, of the invariant strategy by which Nature assigns values to the variables in the distribution. Process features enable us to predict, not merely relationships governed by the underlying distribution, but also how that distribution would CHANGE when conditions are altered, say, by deliberate interventions or by spontaneous transformations.
We will review concepts, principles, and mathematical tools that were found useful in reasoning about causal and counterfactual relations, and will demonstrate their applications in several data-intensive sciences. These include questions of confounding control, policy analysis, misspecification tests, mediation, heterogeneity, selection bias, missing data, and the integration of findings from diverse studies.
©2013 Microsoft Corporation. All rights reserved.