Learning Bayesian network structure from data using integer programming
Dr. James Cussens from the University of York will give a guest talk on Tuesday, December 11th at 14.15 in lecture room 510N3 on the 5th floor in Datablokken Refreshments will be served before the talk.
The structure of a Bayesian network is a directed acyclicgraph (DAG) wich provides compact and intuitive representationof the conditional independence relations between the variables in a joint probability distribution. Bayesian network structure learning BNSL problem becomes one of discrete constrained optimisation.
In this talk will describe how one can use integer programming (IP) to (attempt to) solve this problem. There will be two main foci. On the positie side I will argue the viewing this machine learning problem as constrained optimisation facilitates the explotation of whatever prior knowledge one might have. On the negative side I will address the problem that a naive approach to BNSL with IP requires having too many IP variables - and I will outline ongoing work on what a less naive method would look like.