Distributed Bayesian network structure learning

We propose a new method for learning the structure of a Bayesian network from distributed data sources. Traditional Bayesian network learning takes place at the central site with all data. In many cases, data are distributed over different sites and gathering them at one place is not practical. Our algorithm starts with individual learning at each site with the local data. Then it transmits the learned Bayesian network to the central site. Last, the central site determines the final Bayesian network by looking for frequently occurring parts among the aggregated structures. Experimental results verify that our algorithm successfully finds the same structure that the centralized algorithm produces, with comparable classification accuracy and even higher learning speed.