基于用户评论的图书推荐算法研究

发布时间：2018-05-19 19:44

本文选题：推荐算法 + 协同过滤　；参考：《河北师范大学》2016年硕士论文

【摘要】：互联网的快速发展,尤其是Web 2.0的兴起,为人们提供了丰富大量的信息资源,人们在畅游信息海洋的同时,“信息过载”给人们带来的困惑也越来越多。面对大量的信息,人们往往无从选择,想要寻找自己需要的信息必须花费大量的时间和精力。推荐系统应运而生,很好的解决了信息过载问题,并在电子商务平台得到广泛应用且成为重要组成部分。目前,在众多的推荐技术中,协同过滤是应用最广泛的推荐技术,尤其是在电子商务平台,其应用效果表现的更为突出。在传统的协同过滤推荐系统中,推荐结果的产生是利用用户的评分来完成的。这种方法存在的问题是:一方面,随着用户数和项目数的增加,用户—项目评分矩阵的数据严重稀疏;另一方面,用户的评分反映了用户对所购产品的整体喜好,但用户对产品的某一特征或属性的偏好从整体评分上并不能够得到体现。为了能够充分了解到用户对产品不同特征层面的偏好,大量研究者们通过对用户评论进行特征—情感词对抽取来获取用户偏好,从而为用户提供更准确的推荐。本文针对图书推荐算法,主要从以下几个方面进行了深入的研究和探讨。首先,对用户评论语料进行预处理,抽取出特征—情感词对,量化产品在不同特征层面的分数,构建项目-特征评分矩阵,在此基础上获得用户在项目特征层面的偏好。其次,在进行项目相似度评分预测时,提出利用基于项目的评分相似度和特征相似度的综合相似度来预测评分,填充评分矩阵,解决数据稀疏性问题。然后,针对传统的基于用户的协同过滤算法在用户相似度计算时,只是考虑用户评分上的相似而未考虑用户偏好相似的问题,提出在用户相似度计算时加入偏好相似度计算的方法。最后,使用来自Stanford SNAP的公共图书数据集,通过实验验证本文提出的算法的有效性。实验结果表明,我们的方法与传统的算法相比,达到了良好的推荐效果。
[Abstract]:The rapid development of the Internet, especially the rise of Web 2.0, provides people with a wealth of information resources, while people swim the ocean of information, "information overload" brings people more and more confusion. In the face of a large amount of information, people often have no choice, to find the information they need must spend a lot of time and energy. Recommendation system emerges as the times require, which solves the problem of information overload, and is widely used in e-commerce platform and become an important part. At present, collaborative filtering is the most widely used recommendation technology, especially in e-commerce platform, and its application effect is more prominent. In the traditional collaborative filtering recommendation system, the result of recommendation is produced by the user's score. The problem with this approach is that, on the one hand, as the number of users and items increases, the data of the user-item scoring matrix is severely sparse; on the other hand, the users' ratings reflect the overall preferences of the products they purchase. However, the user's preference for a particular feature or attribute of the product can not be reflected in the overall score. In order to fully understand users' preferences on different feature levels of products, a large number of researchers obtain user preferences by extracting feature-affective word pairs from user comments, thus providing users with more accurate recommendations. In this paper, the book recommendation algorithm, mainly from the following aspects of in-depth research and discussion. Firstly, we preprocess the user comment corpus, extract the feature-affective word pairs, quantify the product scores at different feature levels, construct the item-feature score matrix, and then obtain the user preferences at the item feature level. Secondly, in the prediction of item similarity score, a new method is proposed to predict the score, fill the score matrix, and solve the problem of data sparsity by using the comprehensive similarity based on item score similarity and feature similarity. Then, the traditional collaborative filtering algorithm based on users only considers the similarity of users' scores, but not the similarity of user preferences. This paper presents a method of adding preference similarity to user similarity calculation. Finally, the effectiveness of the proposed algorithm is verified by using the common book data set from Stanford SNAP. The experimental results show that our method achieves a good recommendation effect compared with the traditional algorithm.
【学位授予单位】：河北师范大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【参考文献】