Research On Cognitive Methods In Structure Learning Of Bayesian Networks

Recently, the study of Bayesian networks has become a hot research spot in the field of uncertain Artificial Intelligence (AI), due to its abilities of knowledge presentation and inference. Moreover, Bayesian networks are effective tools for the uncertain decision problems in management. Nowadays, finding new applied usage of Bayesian networks Theories gains much focus, which includes the structure and parameter learning of Bayesian networks and the inference process.Traditional Bayesian networks structure learning methods has two main categories:based on expert-knowledge and datasets. They are processes of determining the network structure by the knowledge comes from mining dependent variables in different patterns of knowledge source. Therefore, Bayesian networks structure learning is considered as knowledge acquiring, which essentially, is studies on Machine Learning. Bayesian networks learning based on expert-knowledge is subjective, and the leaning method based on datasets proved as a NP-hard problem is complicated. Due to the disadvantages of traditional methods, new learning methods are in demand, which make computers has the abilities of human knowledge acquiring and autonomous network structure learning and develop intelligent computing models for learning.This thesis studies Bayesian networks structure learning methods of acquiring and presenting knowledge which are imitating human processes. Based on Bayesian networks theories and new demands in cognitive scientific artificial intelligence fields, we design two knowledge discovery methods in datasets:based on double-bases cooperating mechanism and strong relevant logic, which can autonomous acquiring and presenting knowledge from different patterns of knowledge source. A structure learning method based on multi-value attribute association rules mining algorithm is proposed for complete datasets in problem domain and method based on strong logic is proposed for incomplete datasets. Finally, we prove our methods in the applications of listed corporation finance forewarning and industrial cluster recession modeling.Specifically, the major work includes:(1) Due to the disadvantages of traditional methods and new demands in AI, we analyze the essence of Bayesian networks structure learning, and propose the researching objects and problems. Studies on the framework of Bayesian network theory is conducted, as well as the essence of Bayesian Network structure learning. The two categories of traditional methods are analyzed by the current researching. We also survey the field of cognitive AI.(2) An autonomous cognitive method based on multi-value attribute association rules mining algorithm is proposed for complete datasets in problem domain. First, we design a framework of a series of definitions and theorems which are the basis of cognitive methods. Then several algorithms on knowledge acquiring and presenting in the flows are proposed, including three parts:the method of prior knowledge extraction based on rough sets, kernel principal component analysis and the union of the formers; multi-value attribute association rules mining algorithm based on rough sets; Bayesian network structure presenting method based on causal association rules.(3) The strong relevant logic-Bayesian networks (SRL-BNs) and SRL-BNs based learning methods are proposed. The methods focus on the bottom logic level, and aims to solve the problem of probabilistic logic model in the existing Bayesian networks, and construct the formalization system of Bayesian networks which is based on the strong relevant logic, and defines the components of Bayesian networks by using definitive clause logic language, and gives the declaration semantic. And then the detailed algorithm is proposed, which is able to discover the knowledge autonomously. With the incomplete data set and the data-lacking domain, the algorithm can discover the knowledge autonomously and construct the structure of a Bayesian network. Therefore, it is an automatic machine learning tool of cognitive characteristic. At last, an example is given to illustrate the SRL-BNs components and the modeling procedure.(4) The forewarning model and finance model based on Bayesian networks are studied. Firstly, we studied the survey of background, research significance and the methods in finance forewarning research, secondly we analysis the problems of the existing models. Then propose our Bayesian networks based finance warning model, collect the samples and real data which are used to train the networks, then the parameters and the structure of the Bayesian network can be gained. At last we use the network to predict the warning and analyze the predict results.(5) Based on the strong relevant logic-Bayesian networks, the decay model of industry group is established. We study the status of the decay of industry group. And for the problems which are under the theory study, we use the proposed model to validate them, construct the knowledge base of risk elements of the industry group decay. By the model, we take the real data as the samples to train the model, and to predict the decay probability, and analyze the results.This thesis is aim to simulate the knowledge retrieving process by the learning of Bayesian networks, and propose a series of machine learning methods of cognitive characteristics to construct the Bayesian networks, and then by using the real examples to prove the feasibility of those methods.