Structure Learning Of Bayesian Networks And Its Applications In Data Mining

With the advent of information and communication technologies, there can be no doubt that the cycle of technology development and implementation is accelerating and that we are moving inexorably onward, out of the Industrial Age and into the Information Age. Nowadays, the capture, propagation and exploitation of information are closely related to all of us, information has become perhaps the most important factor determining the standard of living. The society has taken great changes by the innovation of information technology; information estate also provides strong capacity for promoting the development of market economy. Undoubtedly, information will soon have a profound impact on economic growth and society progress in the twenty-first century. The ability of Information management and science decision has enhanced greatly with the establishment of information system, expert system and knowledge base. At the same time, along with the rapid accumulation of data and information, how to use these huge amounts of data is becoming a crucial project for businesses and governments.Bayesian networks are the method for uncertainty reasoning and knowledge representation that was advanced at the end of the 20th Century. It is a probabilistic graphical model, which has been used for probabilistic reasoning in expert systems. Because that the novel method has a powerful ability for reasoning and a flexible mechanism to learning, it provides an effective way to deal with causality or uncertainty. Bayesian networks proved to have surprisingly broad applications, such as medical diagnoses, Image interpretation, pattern recognition, in particular, knowledge discovery and data mining. On this ground, it has been a research hotspot in artificial intelligence and knowledge engineering. In this paper, it focuses on the structure leaning of Bayesian networks, and establishes a systemic theory and method for structure leaning based on the theoretical research and the experimental result. All of these may provide advantageous basis for construction and application of Bayesian networks. To sum up this dissertation, the research and innovation mainly includes as follows.(1) Firstly, this paper discusses the basis theory of Bayesian networks, and then describe the components of it. Moreover, the paper introduce the constitution and semantics of Bayesian network through a sample of Alarm network. It also summarizes the superiorities and characteristics that Bayesian networks compares with other methods. After discussing its function and reasoning mechanism, the paper focus on the aims and major problems of Bayesian networks.(2) Based on the features and attribution, it derives and proves four corollaries of conditional independence with probability parameter. Then used the conditional independence to define the notion of directed acyclic graph (DAG). Discusses the relationship between mutual information with conditional independence, and indicates that we can identify the directly connected variables through computing the structural mutual information for every pair of variables, that is to say, we can identify the direct links in the underlining Bayesian networks. Furthermore, demonstrates the relation between the undirected acyclic graph with dependency model, and studies an approach that tests the conditional independence from contingency tables of the data set.(3) The paper mainly discusses the principle and applications of synergetics, studies the learning mechanism of Bayesian networks structure based on sysnergetics, and establish a systemic approach of structure learning from a new point. It takes Bayesian network structure as a open system, combined with expert knowledge, prior information and observation data, through the cooperative learning of Maximum Aposterior Probability (MAP) and Minimum Description Length (MDL), therefore, the simplicity of the structure and the fitness to the data are naturally leveraged. At last, we can gain a simplest structure model which best fits the given data. All of the work, not only can come over the limitation with subjective partiality, but also avoid to getting into local optimal structure.(4) In section 5, it constructs an effective algorithm for searching the optimal structure from data. First, based on the observation data and expert knowledge, it can learn a undirected acyclic graph which approximate the undirected version of the underlying directed graph by computing the structural mutual information and testing conditional independence. Second, to identify the direction of links with an integrated criterion of MAP and MDL. Eventually, based on the knowledge and rule, to refine the structure based on the global optimization under the integrated criterion, and find the structure model which is best fitness with the given data. Some empirical experiments using two typical databases proved that the algorithm is fairly feasible and efficient.(5) In data mining, it primarily describes the goal, function and technology of data mining, discusses the mainly difficulties in the process. And expounds the notions of data value abundance and data mining benefit evaluation. Furthermore, discusses the framework of data mining, it mainly includes three parts: data preparation, modeling and mining, understanding and evaluation. Meanwhile, it provides the procedure of data mining in every actual step.(6) As for applications, it uses Bayesian networks to geographical information system at the first time, and builds evaluation models of land resource based on Bayesian networks. It can provide the scientific evidences in decision making for utilizing and exploiting land resource, and be very helpful for construction and application expert system in the domain. Finally, the paper summarizes the peculiarities and processes of applying Bayesian networks.As for future work, it should be focus on handling continuous variables and missing values. And continues to study the intelligent learning approach, may use the theory of artificial life such as Cellular Automata (CA) and Ant Behavior. Additionally, to develop a utilitarian software based on the algorithm, which can be used for association, prediction, assorting and clustering in data mining domain.The techniques of data mining will become so successful and so thoroughly integrated into standard database systems that they will no longer be thought of as exotic. Moreover, it is undoubted that Bayesian networks will be a preeminent technology of data mining and get the surprisingly success in the application domains.