Using New Data to Refine a Bayesian Network

We explore the issue of refining an exis­ tent Bayesian network structure using new data which might mention only a subset of the variables. Most previous works have only considered the refinement of the net­ work's conditional probability parameters, and have not addressed the issue of refin­ ing the network's structure. We develop a new approach for refining the network's structure. Our approach is based on the Minimal Description Length (MDL) princi­ ple, and it employs an adapted version of a Bayesian network learning algorithm de­ veloped in our previous work. One of the adaptations required is to modify the previ­ ous algorithm to account for the structure of the existent network. The learning algo­ rithm generates a partial network structure which can then be used to improve the exis­ tent network. We also present experimental evidence demonstrating the effectiveness of our approach.