Abstract: Focusing on the issue that the traditional recommendation system requires users to give a clear preference matrix (U-I matrix) and then uses automation technology to capture the user preferences, a method for mining preference information of Agent from preference database was introduced. From the perspective of knowledge discovery, a k order preference mining algorithm named kPreM was proposed according to the Ceteris Paribus rules (CP rules). The k order CP rules were used to prune the information in the preference database, which decreased the database scanning times and increased the efficiency of mining preference. Then a general graph model named CP-nets (Conditional Preference networks) was used as a tool to reveal that the user preferences can be approximated by the corresponding CP-nets. The theoretical analysis and simulation results show that the user preferences are conditional preferences. In addition, the xcavation of CP-nets preference model provides a theoretical basis for designing personalized recommendation system.

Key words: automation technology, preference database, knowledge discovery, Ceteris Paribus rule (CP rule), qualitative conditional preference network

摘要： 针对传统的推荐系统需要用户给出明确的偏好矩阵（U-I矩阵），进而使用自动化技术来获取用户偏好的问题，提出了一种从偏好数据库中挖掘出Agent的偏好信息的方法。从知识发现的角度，通过Ceteris Paribus规则（CP规则），提出了k阶偏好挖掘算法（kPreM）。在算法中，利用k阶CP规则对偏好数据库中的信息进行剪枝处理，减少了数据库扫描次数，从而提高了偏好信息的挖掘效率。随后以一种通用的图模型——条件偏好网（CP-nets）为工具，揭示了用户的偏好可近似表达为CP-nets的定性条件偏好网。实验结果表明，用户的偏好都是带有条件的偏好。另外，通过挖掘得出的CP-nets偏好模型，为设计个性化的推荐系统提供了理论基础。

关键词: 自动化技术, 偏好数据库, 知识发现, CP规则, 定性条件偏好网