稀疏表达下社会化正则方法与低秩分解推荐模型的研究

项目名称： 稀疏表达下社会化正则方法与低秩分解推荐模型的研究

项目编号： No.71502125

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 管理科学

项目作者： 郁雪

作者单位： 天津大学

项目金额： 18.5万元

中文摘要： 数据缺失下的矩阵低秩分解是协同过滤推荐领域中一个具有挑战性的课题，本项目将用户的社会网络信息引入矩阵低秩分解模型，在研究Web社会网络信息的关联用户信任度问题的基础上，提出了新的社会化正则方法，引导模型学习关联用户空间的潜在低秩结构，理论上具有很好的解释性，为个性化推荐技术研究提供新思路。内容包括：基于社会网络分析的理论挖掘用户之间潜在联系，重新定义关联用户之间信任的描述方法，研究信任传播算法和构建有效的信任测度模型；提出基于社会网络用户信任关系的低秩矩阵分解模型，利用新的先验知识来构造正则项惩罚系数，引导模型向真实评分逼近；研究大规模数据下基于聚类的局部低秩矩阵分解模型，采用遗传聚类方法对原始空间进行划分，在每个子空间上构造新的社会化正则项学习该子空间的潜在低秩结构；最后结合Web社会网络应用进行实证研究，将所研究的方法应用于电子市场行为预测、Web社区推荐和荐和电子商务产品推荐等领域。

中文关键词： 协同过滤；推荐系统；低秩矩阵分解；正则化；社会网络分析

英文摘要： Low-rank matrix approximation in sparse representation is a big challenge in collaborative filtering recommendation area. This project will merge the social network feature data into the low-rank matrix factorization model in which we apply a new regularization method to learning the latent low-rank information within the related user space after exploring the latent relationship between users in social network. The new model aims to provide a new approach on recommendation research with good theoretical explanation. The main research details will include: to explore latent relationship between users based on SNA, redefine the influence factors of trust between related users in social network applications, improve the existed trust propagation algorithm and construct effective trust metric model; develop low-rank matrix factorizations model based on social relation data, propose novel regularizations with prior knowledge in order to handle low-rank approximation with good performance; research cluster-based low-rank matrix factorization in massive data sets, divide the original user space using Genetic-Kmeans algorithm and develop new regularizations to learn the latent structures within each sub-spaces; finally, to complete the empirical study with some real social network applications, and the research model and method will apply to practical areas such as user behavior predicting in electronic market、Web community recommendation and E-commerce item recommendation.

英文关键词： Collaborative Filtering;Recommendation Systems;Low-rank Matrix Factorization;Regularization;Social Network Analysis