Learning the Structure of Related Tasks Alexandru Niculescu-Mizil Department of Computer Science Cornell University Ithaca, NY 14853 alexn@cs.cornell.edu Rich Caruana Department of Computer Science Cornell University Ithaca, NY 14853 caruana@cs.cornell.edu Abstract We consider the problem of learning Bayes Net structures for related tasks. We present a formalism for learning related Bayes Net structures that takes advantage of the similarity between tasks by biasing toward learning similar structures for each task. Heuristic search is used to find a high scoring set of structures (one for each task), where the score for a set of structures is computed in a principled way. Experiments on synthetic problems generated from the ALARM and INSURANCE networks show that learning the structures for related tasks using the proposed method yields better results than learning the structures independently. 1 Introduction Bayes Nets [1] provide a compact, intuitive description of the dependency structure of a domain by using a directed acyclic graph to encode probabilistic dependencies between variables. This intuitive encoding of the dependency structure makes Bayes Nets appealing in expert systems where expert knowledge can be encoded through hand-built dependency graphs. Acquiring expertise from humans, however, is difficult and expensive, so signifi- cant research has focused on learning Bayes Nets from data. The dependency graph also proved to be a very useful data analysis tool. For example Friedman et al. [2] used Bayes Nets learned from gene expression level data to discover regulatory interactions between genes for one type of yeast. In this paper we focus on simultaneously learning Bayes Net structures for multiple prob- lems. Multi-task learning [3, 4, 5] suggests that if different problems are related, learning structures for the problems together should provide an advantage because what is learned for one problem may be useful for the other problems. For example, suppose that gene expression data is available for a number of different species of yeast. It is reasonable to assume that the regulatory interactions between genes for all yeast are similar, but not identical. Learning that there is an interaction between two genes for one species of yeast should provide evidence for interaction between the two genes in other species of yeast. The paper is organized as follows: the next section gives an overview of Bayes Net structure learning. Section 3 describes one approach to simultaneously learning Bayes Nets for multiple data sets. Results of experiments with this method are presented in Section 4.