基于标签关联规则的协同过滤算法研究

发布时间：2018-12-10 07:23

【摘要】：随着互联网的快速普及,信息检索工具的发展经历了三个阶段:从分类导航到搜索引擎,再到现在的推荐系统。推荐系统及相关推荐技术已经不知不觉中深入了人们的生活中,无论是视频网站、音乐网站或APP、社交网站、甚至是平日浏览的新闻网站都离不开推荐技术,处处能看到推荐技术的痕迹。传统的基于协同过滤推荐算法存在很多的缺陷,如稀疏性问题、冷启动问题、可扩展性问题、用户多兴趣问题等等。协同过滤算法只考虑用户间或项目间的相似性来给用户进行推荐,忽略了用户对项目的主观感受。随着Web2.0的发展,在社会化标注系统中加入的标签(TAG)元素为用户提供了一种新的方式来表达对项目的主观感受。标签体现了用户对项目的观点和用户的兴趣,而且也实现了对项目内容相对精确的描述。通过对用户产生的内容(UGC)来对互联网中的户进行社会兴趣挖掘具有非常重要的意义。本文提出了一种引入用户自定义标签内容的基于标签关联规则的协同过滤算法。算法在对评分矩阵填充的过程中引用了基于项目的协同过滤方法,有效的解决了传统的协同过滤算法的稀疏性问题。接着对用户的相似度的计算进行了改进,引入了用户关注度矩阵,对用户评分相似度和用户关注度相似度两部分相似度进行了改进。在这里我们引入Apriori关联规则中计算频繁项集的思想,训练出合适的最小支持度阈值,求出频繁项集,对频繁项集分解得到用户兴趣点,再逆向遍历用户集合,按照用户兴趣点对用户进行聚类。得到用户聚类后,按照前面介绍的改进的用户相似度方法,求出最近邻居用户集合,进而求出用户对项目的预测评分,最后将结果推荐给用户。实验采用MovieLens电影评分数据集,通过一系列的实验对各推荐算法进行对比。实验表明该方法能有效的降低评分矩阵稀疏带来的影响,提高了预测精度。
[Abstract]:With the rapid popularization of the Internet, the development of information retrieval tools has experienced three stages: from classification navigation to search engine, and then to the present recommendation system. Recommendation systems and related recommendation technologies have unconsciously penetrated into people's lives, whether it is video sites, music sites or APP, social networking sites, or even the news sites that they visit all the time are inseparable from the recommendation technology. Everywhere you can see the traces of recommended technology. The traditional collaborative filtering recommendation algorithm has many defects, such as sparse problem, cold start problem, scalability problem, user multi-interest problem and so on. Collaborative filtering algorithm only considers the similarity between users or items to recommend to users, ignoring the subjective feelings of users. With the development of Web2.0, the tag (TAG) element added in the social tagging system provides a new way for users to express their subjective feelings about the project. Tags reflect the user's view and interest in the project, and also achieve a relatively accurate description of the project content. It is of great significance to mine the social interest of the users through the user generated content (UGC). A collaborative filtering algorithm based on tag association rules is proposed in this paper. In the process of filling the scoring matrix, the algorithm uses the item-based collaborative filtering method, which effectively solves the sparse problem of the traditional collaborative filtering algorithm. Then, the user similarity calculation is improved, user attention matrix is introduced, and the user score similarity and user concern similarity are improved. In this paper, we introduce the idea of calculating frequent itemsets in Apriori association rules, train appropriate minimum support threshold, find frequent itemsets, decompose frequent itemsets to get user interest points, and then traverse user sets backwards. Cluster the users according to the points of interest. After the user clustering is obtained, the nearest neighbor user set is obtained according to the improved user similarity method, and then the forecast score of the user is obtained. Finally, the result is recommended to the user. The experiment adopts MovieLens film score data set, and compares each recommendation algorithm through a series of experiments. Experiments show that this method can effectively reduce the impact of sparse scoring matrix and improve the prediction accuracy.
【学位授予单位】：河北工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.3

【参考文献】