Speaker: Peter Spirtes, Carnegie Mellon University
Abstract: Most data-mining algorithms such as neural networks
and decision trees are useful for purposes of classification and diagnosis.
There is another class of important problems for which these algorithms
are not well suited, namely those in which the goal is to predict the effects
of intervening on a system. Answering questions about the effects of interventions
requires knowing the causal relationships among variables.
Bayesian networks consist of a directed graph which represents the causal relationships among variables, and an associated probability distribution. Bayesian networks without hidden variables have a host of desirable properties (e.g. identifiability, finite search space) that make them useful tools for answering questions about the effects of interventions. However, if the goal is to make predictions about the effects of interventions, in most realistic situations the possibility of causes which have not been measured must be considered. While it is possible to represent hidden causes in a Bayesian network, Bayesian networks with hidden variables do not have many of the properties (e.g. identifiability, finite search space) which made Bayesian networks without hidden variable so useful. I will describe a method for dealing with the problem of hidden variables that uses a generalization of Bayesian networks (Partial Ancestral Graphs) which retains many of the desirable properties of Bayesian networks without hidden variables, and a search procedure over Partial Ancestral Graphs which is sometimes feasible even for large numbers of variables. I will also give some examples illustrating Partial Ancestral Graphs and the search procedures, and describe the limitations of current algorithms and some future areas of research.
SHORT BIO: Dr. Peter
Spirtes is Professor of Philosophy and a Professor in the Center
for Automated
Learning and Discovery at Carnegie Mellon University.