Learning Optimal Bayesian Networks - A Shortest Path Finding Perspective

Abstract: In this talk, I will discuss a shortest path finding perspective for learning Bayesian network structures that optimize a scoring function for given data. The main idea is to formulate the problem as finding a shortest path in an implicit state-space search graph. This new formulation raises two orthogonal research issues: the development of search strategies for solving the shortest path finding problem, and the design of admissible heuristic search functions. An A* search algorithm and several data structures for learning provably optimal networks will be presented. I will then describe two heuristic functions which guide the search by relaxing the acyclicity constraint on the problem. Empirical results show that the new formulation outperforms existing state of the art techniques on a variety of benchmark datasets.

Bio: Brandon Malone is a postdoctoral researcher at the University of Helsinki in the Complex Systems Computation Group. Most of his research has addressed improving the scalability of exact Bayesian network structure learning using admissible heuristic search. He has also worked with interdisciplinary groups to investigate problems in epigenetics and metagenomics. He received his PhD from Mississippi State University in 2012.

Host: Sohan Seth