Abstract
We present a block learning algorithm for learning large Bayesian
network structures from limited data in a distributed fashion. Adopting
divideandconquer strategy, firstly, the block learning algorithm
discovers local structures of Bayesian networks individually with
preferred learning techniques. After that, the learned local structures
are combined together to recover the final Bayesian network. A serial of
experiments on golden networks demonstrate that the block learning
algorithm is able to learn a large Bayesian network structure from a
small data set. Moreover, this algorithm is scalable and is capable to
encompass the existing learning techniques. A study on the learning
granularity, the learning engine and the algorithm design discovers that
the block learning algorithm provides a uniform view on various learning
techniques. Consequently, a unifying learning framework is built on the
basis of the block learning algorithm.

Pradeep Ravikumar Last modified: Mon Mar 21 09:30:41 EST 2005