基于用户近邻的上下文张量分解推荐算法

发布时间：2018-05-17 16:04

本文选题：推荐系统 + 矩阵分解　；参考：《江苏大学》2017年硕士论文

【摘要】：计算机的迅速发展带来了严重的信息过载问题,无形之中增加了用户获取自己想要的信息的难度,个性化推荐系统正是在这样的情况下产生了。个性化推荐系统将用户的历史交互行为记录下来并进行详细地分析,基于这些分析的结果为用户推荐他们可能感兴趣的商品。个性化推荐系统不仅为用户带来了较好的用户体验,同时也为网站本身提供了个性化决策机制。本文介绍了个性化推荐系统中所涉及的相关概念、经典的推荐算法、推荐系统的应用场景等,重点研究了二维矩阵分解推荐算法和高维张量分解推荐算法。论文就这些算法存在的数据稀疏性和冷启动问题提出了相应的改进方案。本文的主要工作如下:(1)针对传统的矩阵分解协同过滤算法仍然存在的数据稀疏性问题,提出融合社交信息的矩阵分解推荐模型。通常发现身边的好友的建议会潜移默化地影响我们的购买行为,用户的购买行为不仅跟自己的兴趣相关,同时还会受到他所信任的好友的影响。本文在传统SVD分解模型的基础上,一方面考虑了用户和项目的固有属性对评分的影响,另一方面利用社交网络中的好友关系修正矩阵分解模型,然后使用随机梯度下降法进行矩阵分解。实验结果表明,优化后的算法在实际应用中比传统的SVD推荐算法具有更好的推荐效果。(2)针对基于张量分解的推荐算法存在推荐精度上的问题,提出融合用户近邻信息的N维张量分解算法。首先引入上下文感知信息,把上下文感知中的隐式反馈信息作为张量的第三维度,来建立N维张量分解模型;同时为了进一步提高推荐质量,引入用户近邻信息来优化N维张量分解模型,提高了张量分解推荐算法的准确率。实验结果表明:融合用户近邻的张量分解推荐算法比传统的张量分解算法具有更好的准确性,能有效解决稀疏性和准确性问题。(3)将本文提出的融合社交信息的矩阵分解推荐模型以及融合用户近邻信息的N维张量分解算法应用到软装电子商务系统中,介绍了推荐系统的总体架构和技术选型。该电商系统为了适应多种推荐需求,还使用到了经典的推荐算法。该系统从架构上主要分为数据层、推荐算法层、应用接口层、应用层。该系统的推荐模块主要包括“看了又看”、“买了又买”、“猜你喜欢”、“搭配推荐”等。推荐系统为该电商系统不仅带来了更好的用户体验,同时也吸引了许多的用户。
[Abstract]:The rapid development of computer has brought serious information overload problem, which increases the difficulty for users to obtain the information they want. The personalized recommendation system records the user's historical interaction behavior and analyzes it in detail. Based on the results of these analyses, the users can be recommended for the products they may be interested in. Personalized recommendation system not only brings users a better user experience, but also provides a personalized decision-making mechanism for the website itself. This paper introduces the concepts involved in the personalized recommendation system, the classical recommendation algorithm, the application scenario of the recommendation system, and focuses on the two-dimensional matrix decomposition recommendation algorithm and the high-dimensional Zhang Liang decomposition recommendation algorithm. In this paper, the data sparsity and cold start problem of these algorithms are improved. The main work of this paper is as follows: (1) aiming at the problem of data sparsity still existing in the traditional matrix decomposition and collaborative filtering algorithm, a matrix decomposition recommendation model is proposed to fuse social information. It is often found that the advice of friends around us will affect our purchasing behavior. The user's purchase behavior is not only related to their own interests, but also influenced by their trusted friends. On the basis of the traditional SVD decomposition model, on the one hand, we consider the influence of the inherent attributes of users and items on the score, on the other hand, we use the friend relationship in social network to modify the matrix decomposition model. Then the stochastic gradient descent method is used to decompose the matrix. The experimental results show that the optimized algorithm has better recommendation effect than the traditional SVD recommendation algorithm in practical application. (2) the recommendation algorithm based on Zhang Liang decomposition has the problem of recommendation accuracy. This paper presents a N-dimensional Zhang Liang decomposition algorithm which combines user's nearest neighbor information. Firstly, the context-aware information is introduced, and the implicit feedback information in context-aware is regarded as the third dimension of Zhang Liang to establish the N-dimensional Zhang Liang decomposition model. The user nearest neighbor information is introduced to optimize the N-dimensional Zhang Liang decomposition model, which improves the accuracy of the Zhang Liang decomposition recommendation algorithm. The experimental results show that the proposed Zhang Liang decomposition recommendation algorithm is more accurate than the traditional Zhang Liang decomposition algorithm. It can effectively solve the problem of sparsity and accuracy. It applies the matrix decomposition recommendation model of fusion of social information and the N-dimensional Zhang Liang decomposition algorithm of user's nearest neighbor information to the soft electronic commerce system. The general structure and technology selection of recommendation system are introduced. In order to meet the needs of many kinds of recommendation, the system also uses the classical recommendation algorithm. The system is divided into data layer, recommendation algorithm layer, application interface layer and application layer. The recommendation module of the system mainly includes "read and see", "buy and buy", "guess you like", "match recommendation" and so on. The recommendation system not only brings a better user experience, but also attracts many users.
【学位授予单位】：江苏大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】