The Research On Algorithms For Learning Bayesian Network Structures

Bayesian network is developed by the integration of probability with graph theory. It provides a natural tool for dealing with problem of uncertainty. In recent years, Bayesian network has become a hot research topic in the field of intelligent data processing and has been widely used in expert systems, decision support, pattern recognition, machine learning and data mining. Based on the overview of the research on Bayesian network, this thesis focuses on the research on algorithms for learning Bayesian network structures. The main contents of this thesis are as follows:1. This disertation makes a survey about the research on Bayesian network, including the background, the current research state and development trend of Bayesian network, the basic principle of Bayesian netwrok, the introduction and analysis of the classic algorithms for inferring and learning Bayesian network.2. An algorithm ISOR, based on dependency analysis method, is proposed for learning Bayesian network structure. The algorithm ISOR uses heuristic cut-set searching algorithm and orients all the edges in the network before removing redundant edges. As a result, the algorithm ISOR greatly reduces the number and order of conditional independence tests. Theoretical analysis and experimental results show algorithm ISOR has good performance.3. An algorithm BC-ISOR, based on the underlying principle of Bound &Collapse method and dependency analysis method, is put forward for learning structure of Bayesian network from missing data. The algorithm BC-ISOR first ruturns bounds on the possible estimates of parameters consistent with the available information. The bounds can then be collapded to a point estimate using information about the pattern of missing data. On the basis of parameter estimate, the algorithm ISOR is used to learn Bayesian network structure. Theoretical analysis and experimental results show that the effiencicy of algorithm BC-ISOR is not influenced by the percentage of missing data in the dataset and the algorithm has comparable learning accuracy.