基于张量分解的个性化标签推荐算法

Personalized Tag Recommendation Algorithm Based on Tensor Decomposition

基于互联网的社会标签推荐系统为广大用户提供了一个信息共享平台,让用户以"标签"的形式为其浏览的物品标注信息。标签既描述了物品语义,又反映了用户偏好。标签系统的最大优势在于可以发挥群体的智能,获得用户对物品比较准确的关键词描述,而准确的标签信息是提升个性化推荐系统性能的重要资源。然而,现存的标签推荐系统面临的问题是:由于兴趣的不同,不同的用户对于同一物品可能会打不同的标签,或者是同一标签对于不同用户可能会蕴含不同的语义。因此如何有效获取用户、物品、标签3者之间潜在的语义关联成为标签推荐系统需要解决的主要问题。为此引入三维张量模型,利用三维张量的3个维度来分别描述社会标签推荐系统中3种类型的实体:用户、物品、标签。在基于历史标签数据(标签元数据)构建初始张量的基础上,应用高阶奇异值分解(HOSVD)方法降低张量维度,同时实现3种类型实体之间潜在的语义关联分析,从而进一步提高标签推荐系统的准确性。实验结果表明,该方法较当前两种典型的标签推荐算法(FolkRank和PR)在准确率和召回率性能指标上有明显提升。

Internet-based social tag recommendation system provides a information sharing platform for the majority of users,allowing the users to annotate information for the items they have browsed in the form of＂tag＂.It not only describes the item＇s semantics but also reflects the user＇s preferences.The advantage of tag recommendation system is that the system can play a swarm intelligence to obtain the accurate keywords description of the item,and accurate tag information is an important resource to improve the performance of personalized recommendation system.However,due to the different interests of different users,the existing tag recommendation system is facing aproblem that different users may tag different tags for the same items,and another problem is that the same tag for different users may contain different semantics.So how to effectively get the potential semantic association among the users,items and tags has become a main problem which needs to be solved.Therefore,we introduced a tensor model and used the third-order tensor to describe the three types of entities of social tag recommendation system：users,items and tags.On the basic of constructing initial tensor based on the history tagging data（tagging metadata）,we applied the higher order singular value decomposition（HOSVD）method to reduce the dimension of tensor,at the same time to realize the analysis of potential semantic association between three types of entities.We performed experimental comparison of the method against two tag recommendations algorithms（FolkRank and PR）with two real data sets（Last.fm and Movielens）.Experimental results show significant improvements of the method in terms of recall and precision.