Speaker: Kevin Murphy
University of British Columbia, Departments of Computer Science and Statistics
Friday, April 6, 2007
2:00 pm - 3:00 pm
EBU3b 4140

ABSTRACT
Graphical models are a useful and popular way to represent sparse stochastic relationships between many variables. There is great interest in discovering the structure (topology) of these graphs from flat, unstructured data. However, this is computationally and statistically challenging. In the first part of the talk, I will describe how we use a clever dynamic programming (DP) algorithm (due to Koivisto and Sood) to compute the exact Bayesian posterior over structural features from uncertain interventional data. We apply the method to model some protein flow cytometry data and show that the learned model outperforms previous models. Unfortunately, the DP algorithm suffers from several drawbacks, so in the second part of the talk, I will describe how to overcome these by using it as a proposal distribution for a Metropolis Hastings algorithm. We show that this mixes much faster than traditional proposal distributions. Finally, if there is time, I will discuss some of our work on using L1-penalized logistic regression as a way of performing fast and statistically efficient structure learning in DAG models.

This is joint work with Daniel Eaton and Mark Schmidt.