面向海量高维数据的可深度结合的贝叶斯网学习与推理新方法研究

Bayesian network is a graphical model that based on probability theory,it describes the causal relationships between random variables and represents joint probability distribution efficiently.With the advent of the big data age, the data human acquired are often massive and high dimensional. There are still following problems about learning and inference for complex Bayesian network from massive and high dimensional data: the complexity of traditional learning and inference approaches for high dimensional Bayesian network is too high;existing Bayesian network learning approaches are difficult to handle the situation of the target probability distribution drift over time, and they are difficult to combine prior knowledge; existing researches always separates learning and inference that may result in cost lots of computation to learn a Bayesian network with low inference efficiency .Therefore, this project intends to carry out thorough research on the aforementioned issues, aiming to propose adaptive incremental learning approach for learning Bayesian network, combine prior knowledge or dependency analysis to constrain the search space for leaning Bayesian network, and learn high dimensional Bayesian network based on the strategy of divide and conquer, and propose offline and online exact inference approach for complex Bayesian network; then propose the approach for deep combination of learning and inference in order to learn a Bayesian network with high inference efficiency. The implementation of this project will deepen, expand and promote the research and application of Bayesian network.

贝叶斯网是一种以概率论为基础、描述随机变量间因果关系，并能高效表示全概率分布的图模型，已广泛应用于众多领域。随着大数据时代的到来，人类获得的数据往往存在海量、高维等特点，面向海量、高维数据的复杂贝叶斯网的学习与推理研究还存在如下主要问题：高维数据下传统学习与推理方法复杂度过高；已有学习方法难以有效处理目标概率分布动态变化的情况，且难以结合先验知识；已有方法往往将学习和推理独立开来，可能导致耗费很大计算量却学到推理效率很低的模型。为此，本项目拟围绕以上问题开展研究，提出贝叶斯网自适应增量学习方法、利用先验知识或依赖分析约束搜索空间的方法、基于分治策略学习高维贝叶斯网的方法、面向复杂贝叶斯网的离线和在线精确推理方法，并提出将学习与推理深度结合的方法，使学习过程能够学到易于高效推理的贝叶斯网。以期大幅提高复杂贝叶斯网的学习与推理效率。本项目的实施对深化、拓展贝叶斯网的理论与应用研究具有重要意义。

贝叶斯网将概率论和图论相结合，它是一种描述随机变量间依赖关系、并能紧凑高效的表示联合概率分布的概率图模型，近年来已成为人工智能理论中处理不确定性问题的重要工具.随着大数据时代的到来，人类所获得的数据往往存在海量、高维等特点。本项目紧密围绕面向海量、高维数据的贝叶斯网学习与推理方法中的一些关键问题开展研究。提出具有自适应能力且考虑推理效率的贝叶斯网结构和参数学习新方法，提出贝叶斯网学习过程中结合先验知识的新方法，研究高维贝叶斯网学习中的分治策略，给出面向特定问题的贝叶斯网高效精确推理方法等。并在理论研究基础上开展相关应用研究。. 项目执行期间，发表论文15篇，其中SCI论文9篇，包含6篇CCF-B类或中科院2区论文，被英国女王大学、维也纳工业大学、加州大学河滨分校、法国普罗旺斯微电子研究中心、台湾大学、电子科技大学等的同行引用30余次。申请国家发明专利1项，出版专著2部，获软件著作权1项，获全国商业科技进步二等奖1项。