A recommender system based on tag and time information for social tagging systems

Abstract

Recently, social tagging has become increasingly prevalent on the Internet, which provides an effective way for users to organize, manage, share and search for various kinds of resources. These tagging systems offer lots of useful information, such as tag, an expression of user’s preference towards a certain resource; time, a denotation of user’s interests drift. As information explosion, it is necessary to recommend resources that a user might like. Since collaborative filtering (CF) is aimed to provide personalized services, how to integrate tag and time information in CF to provide better personalized recommendations for social tagging systems becomes a challenging task.

In this paper, we investigate the importance and usefulness of tag and time information when predicting users’ preference and examine how to exploit such information to build an effective resource-recommendation model. We design a recommender system to realize our computational approach. Also, we show empirically using data from a real-world dataset that tag and time information can well express users’ taste and we also show that better performances can be achieved if such information is integrated into CF.

Introduction

With the dramatic development of the Internet, Web 2.0 has emerged and become popular, which transforms users from passive consumers to active producers of content (Zanardi & Capra, 2008). Social tagging systems, such as Del.icio.us (http://delicious.com/), CiteULike (http://www.citeulike.org/), Flickr (http://www.flickr.com/), etc., as typical representatives of Web 2.0, allow users to assign personal labels to resources based on their own background knowledge with a purpose to share, discover and recover resources (Xu, Fu, Mao, & Su, 2006). Along with tagging behaviors, a great deal of valuable information emerged, which strongly suggests the need to make use of such information to provide personalized services.

Among information in social tagging systems, tags and time are two main factors occurred in the process of tagging behaviors. Tags can reflect the interests of a user and as time goes by, users’ lists of tags can be considered as descriptions of the interests they hold (Golder & Huberman, 2006). A tag performs as a bridge between a user and a resource through which user’s preference for the resource is expressed, and the more frequently a tag has been used, the more interested a user is in the related resource. Assuming that Alice often uses “baby health” and “education” as her bookmarks, which show her main interests in the area of baby health and education. Hence, resources tagged with these tags by Alice should be given higher weights than others. Additionally, tags are time sensitive and interest drifts exist in social tagging systems. For instance, Alice previously used plenty of “baby health” to bookmark related resources which denoted her main interest in baby health, probably because she just had a baby to take care of. While at present, she changes her focus to “education” as her child grew up. Thus she concentrates on resources relevant to education. In this case, it is inappropriate to set the same weight over all the resources for Alice. In contrast, a higher weight should be assigned to more recent resources (tagged by “education”) than those appeared to be long time ago (tagged by “baby health”), because more recent bookmarked resources reflect a user’s current interests and they may have strong impacts on future prediction. It becomes necessary to mine and utilize these information from social tagging systems to grasp a user’s current main interests in order to provide better personalized services.

Traditional CF offers a way to provide personalized recommendations. As we discussed above, tag and time information are two important elements in social tagging systems. Therefore, it is necessary to integrate tag and time information in CF to provide effective personalized recommendations for social tagging systems. However, how to exploit tag and time information in a systematic manner in CF remains to be investigated. In this paper, we hypothesize that tag and time information could help improve the quality of personalized recommendations. More specifically, we have built a resource-recommendation model which provides personalized services in social tagging systems by following three phases: rating generation, user similarity calculation and resource recommendation. Under this model, we propose three strategies to generate modified ratings based upon tag and time information. Tag-weight strategy aims to weight each resource based on tag information. Time-weight strategy whose aim is to deal with interest drifts computes weight for each resource based on when a user bookmarked a resource. In tag and time strategy, we discover users’ current main interests by generating rating values simultaneously using tag and time information.

This paper makes the following contributions to the study of personalized recommendations for social tagging systems: (1) it systematically demonstrates that tag information and time information are important when predicting users’ preferences and we develop a computational approach to exploit such information to provide personalized recommendations for social tagging systems, and (2) our proposed recommender system using real data from a social tagging system shows better performances by adding such information in collaborative filtering.

The remainder of this paper is organized as follows: Section 2 contains the literature review. In Section 3, we present details about the computational approach adopted by resource-recommendation model. Section 4 introduces our recommender system. Experimental results and analysis are given in Section 5; and we conclude and discuss our future work in Section 6.