基于协同过滤的评分预测算法研究

发布时间：2019-01-03 09:30

【摘要】：随着互联网中用户、商品、交易记录、社交信息等一系列数据的爆炸式增长,海量规模的信息资源充斥在网络中,容易产生"信息过载"现象。为解决这一问题,个性化推荐技术应运而生,它能够为用户提供符合其自身特性的信息服务和决策支持。协同过滤算法是个性化推荐技术中的热门研究课题,其通过分析用户行为,在用户群体中挖掘其他与指定用户兴趣相仿的用户,综合这些相似用户特征对目标物品进行评分,形成推荐模块对目标物品的评分预测。然而,随着数据规模的不断增长,协同过滤推荐算法也面临一系列挑战,诸如数据稀疏性问题、扩展性问题、推荐准确性问题等。本文深入研究基于模型的协同过滤推荐算法中的数据稀疏性、可扩展性问题,并对基于受限玻尔兹曼机和基于矩阵奇异值分解的评分预测推荐算法进行改进。主要工作包括:第一,对传统协同过滤推荐算法的体系结构、发展现状展开研究,对基于邻居相似度和基于模型的协同过滤推荐算法分别作详细介绍。深入研究RBM模型的网络结构、对比散度训练方法,详细分析了奇异值分解(Singular Value Decomposition,简记为SVD)模型的理论方法,对隐语义模型及正则化方法进行了详细说明。第二,对基于RBM的协同过滤推荐算法进行改进,加入训练数据中用户浏览过但未评分的行为信息,形成基于条件受限玻尔兹曼机(Conditional RBM,简记为CRBM)的协同过滤预测算法,并对现有的针对用户的CRBM进行改进,提出针对项目的CRBM模型。实验结果表明,改进后针对项目的CRBM算法预测精度优于目前针对用户的CRBM协同过滤算法。第三,对基于用户行为属性的SVD++预测模型进行分析与改进,加入用户历史行为记录的潜在信息,用包含用户喜好的隐性特征向量矩阵替换原SVD模型中的用户特征向量矩阵,提出非对称奇异值分解算法(Asymmetric SVD,简记为ASVD)及其对偶模型,并对提出的预测模型进行扩展,加入k近邻关系形成融合推荐模型进行评分预测。实验结果表明,所提出的融合模型能有效提高推荐系统的预测精度。
[Abstract]:With the explosive growth of a series of data such as users, commodities, transaction records, social information and so on in the Internet, massive information resources are flooded in the network, which is easy to produce the phenomenon of "information overload". In order to solve this problem, personalized recommendation technology emerged as the times require, which can provide users with information services and decision support according to their own characteristics. Collaborative filtering algorithm is a hot research topic in personalized recommendation technology. By analyzing user behavior, the collaborative filtering algorithm can mine other users with similar interests to specific users and score the target items by synthesizing these similar user characteristics. Form recommendation module to predict target item score. However, with the increasing of data scale, collaborative filtering recommendation algorithms face a series of challenges, such as data sparsity problem, scalability problem, recommendation accuracy problem and so on. In this paper, the data sparsity and extensibility of model-based collaborative filtering recommendation algorithm are studied in depth, and the evaluation prediction recommendation algorithm based on constrained Boltzmann machine and matrix singular value decomposition is improved. The main works are as follows: first, the architecture and development status of traditional collaborative filtering recommendation algorithms are studied, and the neighbor similarity and model-based collaborative filtering recommendation algorithms are introduced in detail. This paper deeply studies the network structure of RBM model, compares the divergence training method, analyzes in detail the theoretical method of singular value decomposition (Singular Value Decomposition,) model as SVD), and explains the implicit semantic model and regularization method in detail. Secondly, the collaborative filtering recommendation algorithm based on RBM is improved, and the behavior information that the user browses but does not score in the training data is added to form a collaborative filtering prediction algorithm based on conditional constrained Boltzmann machine (Conditional RBM, abbreviated as CRBM). The existing CRBM for users is improved, and the CRBM model for the project is proposed. The experimental results show that the prediction accuracy of the improved CRBM algorithm is better than that of the current CRBM collaborative filtering algorithm for users. Thirdly, the SVD prediction model based on user behavior attributes is analyzed and improved, and the latent information of user history behavior record is added to replace the user eigenvector matrix in the original SVD model with a recessive eigenvector matrix containing user preferences. An asymmetric singular value decomposition (Asymmetric SVD,) algorithm called ASVD) and its dual model are proposed. The proposed prediction model is extended and the k-nearest neighbor relationship is added to form the fusion recommendation model. Experimental results show that the proposed fusion model can effectively improve the prediction accuracy of the recommendation system.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】