Learning Bayesian Networks with Local Structure



AUTHOR: Nir Friedman

AFFILIATION: Stanford University

ABSTRACT:

{\em Bayesian networks\/} are arguably the method of choice in artificial intelligence for representing and reasoning with probabilistic knowledge. These networks exploit the structure of the domain (in terms of conditional independences) to allow compact representation of probability measures and efficient inference procedures, and are currently used in numerous applications. A major problem in practice is the aquisition of these networks, which can be often expansive. Thus, there is a growing body of work on learning Bayesian networks from raw data, which is often readily available. I will start the talk with a quick overview of our current project on learning probabilistic models from data.

I will then focus on a novel extension to the known methods for learning Bayesian networks from data that improves the quality of the learned models. Current methods learn a Bayesian network by searching the space of network structures. The quantification of parameters in each candidate structure is done by estimating the full {\em conditional probability table\/} (CPT) for each variable. Our approach explicitly represents and learns the {\em local structure\/} in the CPTs. This increases the space of possible models, enabling the representation of CPTs with a variable number of parameters that depends on the learned local structures. The benefit is that the learning routine is capable of inducing models that emulate better the real complexity of the interactions observed in the data.

I will describe the theoretical foundation and practical aspects of learning local structures, as well as an empirical evaluation of the proposed method. This evaluation compare this approach with the standard learning procedure. The resulting learning curves of the procedure that exploits the local structure converge faster than these of the standard procedure. Our results also show that networks learned with local structure tend to be more complex (in terms of arcs), yet maintain less parameters.

This talk describes joint work with Moises Goldszmidt of Rockwell Science Center.