The Study Of Learning Bayesian Network With Missing Values

This thesis is about study on learning Bayesian network with missing values, andthereby, presents several creative and e?cient solutions.Firstly, new SEM-PACOB and SGS-PACOB algorithm is presented based on thevaluable network structure candidates found by a parallel ACOB solution. Experi-ments show that, the two algorithms make attractive improvements on the quality ofthe learned Bayesian networks.Secondly, a DSEM-PACOB algorithm, an e?cient approach based on SEM algo-rithm, is proposed. This algorithm uses rational tactics to choose several node vari-ables, which have close correlations with the estimated node variable, and to estimatemissing values. The experiments show that, comparing with SEM and SEM-PACOBalgorithm, DSEM-PACOB algorithm gets qualitative improvements on the quality ofthe final learned Bayesian network, and also makes the algorithm converge to idealresults quite smoothly.Thirdly, mainly serving for SEM-PACOB and DSEM-PACOB algorithm, how toselect an initial network is studied. Su?cient experimental results show that the initialnetwork plays a important role on convergence speed of the learning algorithms. It isconcluded that if the dataset'scale is small and missing values probability is high, thecloser matching between initial network and training dataset, the better final resultsare.Last, a hybrid heuristic idea is introduced to DSEM-GS-PACOB algorithm. Thealgorithm skillfully uses characteristics of ACO algorithm, and combines two statisticinformation of the SGS-PACOB and DSEM-PACOB algorithm as its PACOB algo-rithm's heuristic information. The experiments show DSEM-GS-PACOB algorithmfully out-performs both SGS-PACOB and DSEM-PACOB algorithm, and makes thealgorithm converge to ideal results smoothly. Comparing with those algorithms hav-ing only one data completion policy, DSEM-GS-PACOB algorithm not only achieves astable Logloss value, and also makes improvements on the learned Bayesian networkstructure.