偏好建模、获取、推理，是计算机科学中涉及多学科交叉的研究领域，已成为人工智能研究中的一个基本主题，在推荐系统、社交网络、国家安全、产品设计与营销、个性化定制等方面具有相当重要的学术价值和十分现实的应用价值。条件偏好网(Conditional Preference networks，CP-nets)是用来表示偏好的主要图形模型之一，能够简洁地表示属性或者变量之间的定性偏好关系。 CP-nets已被深入研究并应用于涉及偏好处理的各种问题，包括表示、推理、聚合和学习。其中，CP-nets的学习问题更为引人关注，并取得了一系列的成果，如对不一致数据的学习、含有噪声数据的学习、近似学习等。然而，CP-nets学习相关的理论尚未完善，重点是动态和多属性的CP-nets学习问题、多Agents的偏好学习方法有待深入研究。针对现有研究工作存在的不足和问题，本文从CP-nets的学习基本方法出发，研究CP-nets的结构学习问题，其主要研究工作和创新成果如下。 针对CP-nets结构学习中存在的数据形式、属性父集判定和无环学习等关键共性问题，深入研究CP-nets结构学习的基本方法。首先给出偏好数据库设计，将常见的数据集转化成偏好数据库的形式；其次基于条件偏好的判定，给出条件属性父集求解的判定和支持度计算，设计基于最小描述长度的评分函数；最后，给出CP-nets结构学习方法，包括结构学习算法、去环算法、条件偏好表求解以及算法分析。在模拟数据和真实数据集上的实验及对比试验表明，本文提出学习方法具有较高的相似度、相容度。 针对CP-nets结构变化数目随着属性数量呈指数形式增长，而无法采用遍历法学习的难题，采用A*算法学习CP-nets结构，以近似解逼近其最优结构。将CP-nets的结构学习问题看作利用给定数据集的得分函数求取最短路径问题。对启发式函数进行约束，将无环目标加入启发式函数，改进现有的启发式函数。用状态空间搜索图表示学习问题的解空间，以所得图中的起始结点和目标结点之间的最短路径为参考，直接得到无环CP-nets结构。本文方法将CP-nets结构学习的时间降低到多项式时间。实验表明，在几乎不影响相似度和相容度的前提下，缩短了多属性CP-nets结构学习的计算时间。 针对用户偏好通常会随时间变化而变化，偏好数据以数据流的形式存在的现状，本文突破了现有CP-nets结构学习聚焦于静态场景、用户偏好信息固定不变的学习模式，设计基于滑动窗口建立流式偏好数据库模型，给出偏好数据流的增量式CP-nets结构学习方法，尝试解决流式偏好数据的学习问题。在模拟数据和真实数据集上的实验结果表明，增量式方法学习CP-nets对动态数据的结构学习的有效性。 针对多Agents聚合学习问题，本文提出了基于概率CP-nets(PCP-nets)的多Agents学习框架，即在学习单个Agent的偏好模型基础上，利用Max聚合的方法，学习得到PCP-nets，表达多个Agents的偏好。在模拟数据和真实数据集上的实验结果表明，聚合方法学习可得到有效的CP-nets结构，并具有较高的相似度和相容度，是对多Agents聚合学习的有益探索。此外，提出针对多Agent的并行化CP-nets学习研究思路，对分布式计算具有借鉴意义。 综上所述，本文针对CP-nets的结构学习问题，在基本方法和框架、多属性结构学习、流式数据结构学习、多CP-nets结构学习等方面取得了新颖的研究成果，并在不同的数据集上进行了验证，与其他方法进行了对比。最后，本文探讨了CP-nets结构学习今后的研究方向，包括采用属性约简、深度学习和贝叶斯网络迁移学习等方法和技术在CP-nets结构学习中的应用。 关键词：人工智能；机器学习；条件偏好网；偏好数据库；结构学习；A*算法；数据流；增量式方法；聚合

 Modeling, learning and reasoning of preference are the research fields involved in multidisciplinary research in computer science. They have become a basic theme in artificial intelligence. Moreover, they have considerable academic value and practical value which can be used in recommendation systems, social networks, national security, design and marketing of product, personalized customization, etc. One of the main graphical models used to represent preferences is Conditional Preference Networks (CP-nets), which can succinctly represent qualitative preferences among attributes. CP-nets have been thoroughly researched and applied to a variety of issues involving preference processing which including representation, reasoning, aggregation, and learning. Among of them, the learning problems of the CP-nets have attracted more attention and have achieved a series of results, such as learning from inconsistent data, learning with noise data, and approximate learning of optimal results. However, the theory of CP-nets learning has not been perfected. In particular, the learning problem such as large-scale attributes, dynamic data and the multi-agents preference need to be further studied. In view of the problems existing in the previous research work, this thesis starts from the basic methods of CP-nets learning, and studies the structural learning of CP-nets, mainly carries out the following research work. Aiming at the key problems of data form, determination of parent set of attribute and acyclic structure learning of CP-nets, this thesis studies the basic structure learning methods of CP-nets. Firstly, the design of preference database is introduced, and the common data set can be transformed into the form of preference database. Secondly, the judgment and support degree of the conditional attribute parent set solution is given based on conditional preference, and the scoring function is designed based on the minimum description length. Finally, the structure learning of CP-nets is provided based on the above mentioned scoring function, including the structure learning algorithms, acyclic algorithms, conditional preference table algorithms, and algorithms analysis. The experimental results on the simulated data and the real data set demonstrate the higher performance of the proposed method. Aiming at the large number of attributes in the CP-nets, the number of structure changes increases exponentially, and the structure of CP-nets cannot be learned by traversing method. The A* algorithm is used to learn the structure of CP-nets and approximate its optimal structure. In this thesis, the structure learning problem of CP-nets is considered as the problem of finding the shortest path by using the scoring function for a given data set. The acyclic object is added by constraining and improving the heuristic function. The solution space of the learning problem is represented by the state space search graph. The shortest path between the starting node and the target node is taken as a reference, and the acyclic CP-nets structure can be obtained directly. In this thesis, the time of structure learning of CP-nets is reduced to polynomial time, and the experimental results show the proposed method can shorten the calculation time with affecting the similarity and agreement obviously. In view of the fact that user preference usually varies with time and preference data exists in the form of data stream, the thesis changes the existing structure learning of CP-nets which focuses on static scenes, and designs a stream preference database model based on sliding window, then gives the incremental CP-nets structure learning method based on the preference data stream to solve the learning problem of streaming preference data. The experimental results on the simulated data and the real data set show that the incremental method learning CP-nets is effective for data stream. In order to learn preference of multi-agents, this thesis proposes the learning framework based on probabilistic CP-nets (PCP-nets). The thesis takes advantage of Max aggregation method to learn PCP-nets and express preference of multiple agents upon learning of a single agent's preference. The experimental results on the simulated data and the real data set show that the aggregation method of learning can obtain effective CP-nets structure, and has high similarity and compatibility. It is a useful exploration for multi-agents aggregation learning. In addition, the research idea of parallelized CP-nets for multi-agent is proposed, which is of great significance for distributed computing. In summary, this thesis has obtained relevant research results for the structural learning problems of CP-nets in basic methods and frameworks, structure learning of large-scale attributes, structure learning of data stream, and structure learning of multiple CP-nets. The above mentioned methods are verified in different data. Finally, the thesis explores the future research directions of structure learning of CP-nets, including methods and techniques such as attribute reduction, deep learning and bayesian network migration learning. Keywords: Artificial intelligence, machine learning, conditional preference networks, preference database, structure learning, A* algorithm, data stream, incremental learning, aggregation.