Bayesian Network Structure Learning And Application

Since the beginning of the 21st century,the field of artificial intelligence has made great progress.As an graphic model for representing uncertainty knowledge,bayesian network(BN)has become one of the hotspots and important achieve-ments in the field of artificial intelligence.Buliding the corresponding structure of bayesian network is the fundamental part of applying and reasoning bayesian networks,and the traditional expert knowledge based method has been gradually replaced by the structure learning methods.However,the number of possible structures increases exponentially with the number of nodes,which cause the structure learning methods suffer the problems as low efficency or slow conver-gence speed,and the structure learned by these methods always differs from the practical one.Therefore,it is of great theoretical and practical significance to study the bayesian network structure learning problem.In this dissertation,we mainly study the bayesian network structure learning problem,and also consider its application in multi-label classification.Consider-ing the size of the newtork nodes,this dissertation proposed two kinds of structure learning methods,and further apply the bayesian network to the multi-lable clas-sification problem.The main works and contributions of this dissertation are listed in the following parts.Firstly,consider that the search-and-scoring based structure learning meth-ods always suffer the problems as easy trapped in local optimal or low convergence speed,this dissertation proposes a modified artificial bee colony-based bayesian network structure learning algorithm.The directed acyclic graph structure is transferred to the food souce in the artifical bee process,which turansforms the structure learning problem into the searching for otpimal food source problem.The mutation and crossover operators of differcnctial evolution are modified into the neighbourhood searching mechanism.In order to ensure the feasible of the structure in the searching process,an acyclic correct procedure based on depth-first-serach is proposed.The Markov chain of the food source is establised to prove that the proposed algorithm converges to global optimal with probability 1.Experimental results manifest that the proposed algorithm gains better structure scoring and convergence speed.Secondly,consider that when the node size is very large,the search space would be very huge,which cause the complexity of the search-and-scoring based methods be very high,this dissertation proposes a new structure learning method that hybrid the constrain-based with the search-and scoring based method.The conditional independence test is used to learn the unstructured independent graph structure.Then the undirected independent graph is decomposed by the idea of recursive decomposition to construct a number of subset undirected graph struc-tures,and the edges and directions of the subset structures are learned by the serach-and-scoring based method.Finally,the sub-structure synthesis rules are used to reconstruct these substructures into the required directed acyclic graph structures.Compared with the search-and-scoring based methods,the hybrid method is more effective when learning the bayesian network structure with large number of nodes.Lastly,consider that the labels in the multi-label classification problem are always dependent with other labels,this dissertation presents a bayesian network-based multi-label classification approach.Each node in the bayesian network represents a label,and the arcs and conditional probability captures the depen-dency of the labels.The structure of the network structure is learned from the trainning data,and we use the maximum likelihood estimate method to learn the parameters of the structure.For each test set,selects its adjacent sample sets by the Euclidean distance in the traning sample,and the label information of the adjacent sets are inputed into the bayesian network as evidence.The junction tree algorithm is used to predict for the possible labels of the test set.