基于知识迁移的跨领域推荐算法研究

发布时间：2018-02-14 00:56

本文关键词： 知识迁移跨领域推荐用户兴趣度知识模型　出处：《北京交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：在当今的互联网时代下,大量信息数据的积累使得人们很难迅速准确地发现自己所感兴趣的内容,推荐系统在一定程度上解决了这种信息过载问题,但是传统的推荐系统难以解决冷启动、数据稀疏性问题。随着互联网的普及,不同领域的信息可以共享和互为补充,为解决冷启动问题,跨领域推荐带来了机遇。为提高跨领域推荐结果的准确性和多样性,提升跨领域信息资源的利用率,本文提出了两种基于领域间知识迁移的跨领域推荐算法。主要工作如下:(1)本文首先对源领域和目标领域的用户进行分析,在领域间有用户重叠的场景下,提出了基于用户兴趣相似度迁移的跨领域推荐算法(User interest-based transfer,UIT);从用户角度出发,用户的兴趣度会在不同领域中体现,用户的好友在不同领域中不尽相同。基于此,可通过用户的好友将用户在源领域中的兴趣度迁移到目标推荐领域中去,我们首先把源领域的信息评分矩阵进行填充,再利用矩阵分解方法计算用户兴趣度,最终我们得到源领域中的兴趣度与目标域中改进相似度的融合算法。(2)针对领域间没有用户重叠的场景,我们进一步提出了基于共享知识模型的跨领域推荐算法(Sharing knowledge pattern,SKP),通过分析各个领域中用户-项目-评分数据,可以得到用户的潜在特征和项目的潜在特征,在将用户和项目的潜在特征分别聚类的基础上,得到用户分组对项目分组的评分知识模型,最终充分利用目标领域的个性知识模型和共享各个领域的共性知识模型来提供最终的推荐结果。(3)在Spark集群环境下,我们对本文提出的算法以及相关对比算法进行了并行化实现和优化。实验结果表明,与单一领域的协同过滤算法和目前的跨领域算法相比,本文提出的算法有较低的RMSE,较高的准确率、召回率和F1值,并且在不同的Spark集群节点数量下验证了本文提出的算法更具可扩展性和实时性。
[Abstract]:In the current Internet era, the accumulation of a large amount of information data makes it difficult for people to quickly and accurately find the content they are interested in. The recommendation system solves the problem of this kind of information overload to a certain extent. However, the traditional recommendation system is difficult to solve the problem of cold start and data sparsity. With the popularization of the Internet, information in different fields can be shared and supplemented each other, in order to solve the cold start problem, Cross-domain recommendation brings opportunities. In order to improve the accuracy and diversity of cross-domain recommendation results, improve the utilization of cross-domain information resources, In this paper, two cross-domain recommendation algorithms based on inter-domain knowledge migration are proposed. In this paper, a cross-domain recommendation algorithm based on user interest similarity migration is proposed. From the user's point of view, the user's interest will be reflected in different fields, and the user's friends will be different in different fields. The user's interest in the source domain can be transferred to the target recommendation field through the friends of the user. We first fill in the information scoring matrix of the source domain, and then calculate the user's interest by using matrix decomposition method. Finally, we get the fusion algorithm of interest degree in source domain and improved similarity degree in target domain. We further propose a cross-domain recommendation algorithm based on shared knowledge model, which is sharing knowledge patternSKP. By analyzing the user-project-score data in each domain, we can get the potential features of users and the potential features of projects. On the basis of clustering the potential features of users and projects, the scoring knowledge model of user groups for project grouping is obtained. Finally, we make full use of the individual knowledge model of the target domain and the common knowledge model of each domain to provide the final recommendation result. (3) in the Spark cluster environment, We parallelize and optimize the proposed algorithm and the correlation contrast algorithm. The experimental results show that, compared with the single domain collaborative filtering algorithm and the current cross-domain algorithm, The proposed algorithm has lower RMSE, higher accuracy, recall rate and F1 value, and it is verified that the proposed algorithm is more scalable and real-time under different number of Spark cluster nodes.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】