CSE459 Causal Discovery (2003)

Lecturer: Kevin Korb

Lectures: Wednesday 10.00-12.00 (M3/Bld13A).

Office hours: Monday 4.00-5.00pm, Tuesday 2.00-3.00pm or by appointment.

Outline of Causal Discovery

Causal discovery aims to develop algorithms to learn the structure of causal processes from observation. This goes to the heart of the long-standing dispute over whether we can learn causal relations from observed correlations. We will start with causal modeling techniques introduced in the early 20th century by Sewall Wright for dealing with linear causal structure. We will briefly review developments throughout the last century dealing with linear models and their causal interpretation, including structural equation modeling in economics.

In the late 1980s graphical models - Bayesian networks - became popular in AI for representing and reasoning with probabilities. In order to overcome the "knowledge bottleneck", researchers quickly turned to the problem of the machine learning of Bayesian networks from data. The techniques discovered are natural extensions of the linear modeling above. We will examine the main developments: The Verma-Pearl Conditional Independence Algorithm (1990), Tetrad II's PC Algorithm (1993), Cooper-Herskovits's Bayesian K2 (1991), Heckerman and Geiger's BDe/BGe (1995), Causal Discovery via MML (1996).

We will consider the question whether Bayesian networks are properly understood as fundamentally causal or simply probabilistic (i.e., correlation vs cause, again).

We also look at closely related questions, such as: learning probabilities from data; learning with incomplete data; Monte Carlo methods for automated learning; expectation maximization (EM) methods; evaluating machine learning methods.

The exact composition of these topics and their relative weights in the schedule may vary.

Prerequisites

The module ``Bayesian Models'' is recommended.

Lectures

Lectures will be 2hr per week: 10.00-12.00noon; (M3/Bld13A); [Wks10-15]

Lecture notes:

Lecture 1 Machine Learning
Lecture 2 Bayesian networks
Lecture 3 Bayesian Confirmation Theory
Lecture 4 Linear Causal Models
Lecture 6 CI Learning
Lecture 7 Parameter Learning
Lecture 8 Metric Learning

Assessment

Assessment will be in the form of a presentation and short paper. The paper is due 27 June.

Some possible paper topics.

Readings for lectures

Readings will be handed out. Some useful books are:

J Pearl (2000) Causality.
R Neapolitan (2003) Learning Bayesian networks.
Michael I. Jordan (Ed.) (1998) Learning in graphical models. MIT.
Peter Spirtes, Clark Glymour, Richard Scheines (2000) Causation, prediction, and search. Springer Verlag.

Interesting WWW Pages

Kevin Korb