Research On The Discovery Algorithm Of Maximum Ancestral Graph Structure Based On Markov Blanket

Markov blanket discovery is often used for causal feature selection.The theory proves that given the Markov blanket variable set of the target variable,the addition of other variables cannot add additional information to the target variable.As a local variable set with special meaning in the causal diagram,the Markov blanket can be combined with it to outline the causal structure through local expansion,which will not only make the result more accurate but also more meaningful.The most widely studied type of causal graph is the directed acyclic graph,but when latent variables and selection bias appear in the system,it cannot correctly represent all independence relations,and a more reasonable type of causal graph is the maximum an-cestral graph.Based on the advanced nature of Markov blankets,this paper takes the discovery of Markov blankets as the starting point,and conducts research on the local structure discovery and global structure discovery of the maximum ancestral graph.The specific results are as follows:First,for the discovery of the local structure of the maximum ancestral graph,this paper proposes a Markov blanket-based learning method MBLC.By sequentially learn-ing the target variable and the Markov blanket of its Markov blanket,the local structure and direction are continuously expanded until The direction of the neighbors of the target variable cannot be further determined.Second,for the discovery of the global structure of the maximum ancestral graph,this paper proposes a Markov blanket-based learning method MBSL.In the process of travers-ing the Markov blanket of each variable,the final global structure is obtained through local splicing.And use methods such as interior point jumping to speed up construction effi-ciency.Through simulated sampling experiments,the method proposed in this paper is com-pared with the classic maximum ancestral graph discovery algorithm FCI.The experi-ments show that the method proposed in this paper is more efficient and accurate in the structure discovery results,not only the number of conditional independent tests is small,but also in the edge And the discovery of arrows will be more accurate.Finally,the two methods proposed in this paper are applied to the financial data of treasury bonds at the same time to find the surrounding structure of the ten-year treasury bond yield.Practice has proved that both methods can identify the same causal relation-ship,and the causal and influencing variables obtained based on this have been verified in both correlation and causality.