The Research On Intervention Learning Of Bayesian Networks Based On Sensitivity Analysis

The discovery causal relationship of structural model is an important research content in area of machine learning. The causal discovery methods of structural model are two: One is passive learning based on observational data, and the other is active learning based on observational data and experimental data. Based on the two approaches, theory of sensitivity analysis is introduced, and features of causal structure and parameter are discussed based on causal structural model, and character analysis of causal structural model is effective by experimental verification .The main work of this thesis are as follows:(1)This thesis makes a survey about current situation on structure learning, parameter learning, intervention learning and sensitivity analysis.(2) The causal relationships of Bayesian networks from observational data are difficult to efficiently be discovered, so the learning methods of combination observational data and intervention data are used for learning causal networks. As interventional node is difficult to be determined in the active learning setting, a active learning algorithm in the causal network is presented for selecting intervention node based on local sensitivity analysis (ILPSA); The interventional node is selected by local network parameters sensitivity analysis, and the interventional node is manipulated to occur the intervention data; Then, the causal network is learned by MLE from combination observational data and intervention data, and the learning networks results of learning are measured by KL divergence. Experimental results show that ILPSA algorithm is better than interventional node with choosing randomly and MLE of passive learning from small samples.(3)The causal structure can not be learned completely with observational data, we have to collect further information on causal structures from experiments with external interventions. An active learning method (Intervention Learning of Local Sensitivity Analysis–ILLSA)is proposed for discovering original causal structure from observational data, and this causal casual structure is decomposed into many cliques based on junction tree construction algorithm, then the most important edges of every clique are found by sensitivity analysis. We orient the undirected edges via intervention experiments separately. In the experiments, some variables are manipulated through external interventions-sequencial interventions. Experimental results show that method of our active learning is better than interventional node with choosing randomly and passive learning method.