关键词: 社会标签,标签推荐,张量分解,高阶奇异值分解(HOSVD)

Abstract: 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 a problem 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.

Key words: Social tag,Tag recommendation,Tensor decomposition,Higher order singular value decomposition