Learning Bayesian Networks Through Knowledge Reduction

Student thesis: Master thesis (including HD thesis)

  • Jorge Pablo Cordero Hernandez
Learning Bayesian networks from data becomes intractable when a large number of variables are involved in the application domain. Much effort has been made in the past to overcome the computational problem using the divide and conquer strategy. In this Master thesis, it is provided a prior solution to this strategy by introducing a general class of models, named the Bayesian network knots, which explicitly partition the variables into several local components in the network. We propose a learning algorithm called the Overlapping Expansion Learning (OSL) algorithm. Furthermore, we investigate the implications of attribute clustering for learning Bayesian networks. Experimental results show that the OSL is highly competitive. Moreover, we developed a novel attribute clustering algorithm, named the Star Discovery (SD) algorithm. The SD algorithm is able to discover groups of variables with a higher performance than several attribute clustering approaches.
Publication dateAug 2007
ID: 61071122