Knowledge learning and reasoning with uncertainty is the basis of intelligent behavior. The explosive growth of the amount of data caused by the Internet revolution make us enter the era of big data. The data has become an important resource for various application areas,showing enormous volume, fast changing, diversity, structure complexity. The purpose of analyzing the data is to integrate and effectively present the information associated, is to predict the future. Effective analysis and mining on these large, fast changing and diverse data with complex structures, has become the hottest direction of research in today's technology sector, but also the support of the key technologies of the data industry. Research on the key technologies of uncertainty of knowledge representation and reasoning from the vast amounts of data, has important theoretical significance and practical value. Bayesian network is a tool for uncertain knowledge learning, reasoning and data analysis, which brings probability and statistics into the fields of complex system. Despite its solid theoretical foundation, in the fields of knowledge synthesis and reasoning mechanism with uncertain for Bayesian network ,there are still many problems to be further explored. Due to the characteristics of the model itself, the problems of structure learning and inference are proved to be NP-hard. In this paper, we do theoretical analysis on three aspects: building essential graph, triangulation of the moral graph, sampling algorithm for approximate reasoning, and put forward a series of effective algorithms. The mainly work and contribution of this thesis are as follows: Firstly, a new BN structure-learning algorithm based on Maximal Information Coefficient and conditional independence tests, is proposed. Starting from an graph structure without any edges, our proposed method constructs an undirected graph structure based on measure the dependency between two variables using Maximal Information Coefficient. Next, these edges are oriented direction based on conditional independence test, which generates a draft DAG. Then, we greedily explore the optimal structure in the space of equivalence classes using the essential graph of draft DAG as a seed structure. The algorithm was tested on four standard network structures. The experimental results explain that our method can efficiently and accurately identify BN structures from data. Compared with the original algorithm, our proposed alg...
修改评论