Article Preview
Top1. Introduction
Folksonomies, also known as collaborative tagging systems, allow users to annotate, share and manage their resources with tags. Delicious (http://www.citeulike.org/) are two popular websites based on folksonomies which support users to collect different resources such as URLs or publications according to their preference and specify the chosen resources with some words namely tags. However, with the explosive growth of the Internet, finding the relevant or interesting resources from thousands, even millions of websites is a challenge for users when they involve in folksonomies.
Personalized recommendation technique is proved to be one of effective ways to solve the above problem. The task of a recommendation system is to recommend the resources that satisfy the user’s interest based on the historical records of his/her activities. Here resources could be photos, movies, URLs, publication references, etc. For example, a movie recommendation system recommends the movies according to user’s preference on the ones he/she has watched. Classical recommendation methods, which can be broadly classified into collaborative filtering, content-based (Adomavicius & Tuzhilin, 2005), and diffusion-based (Zhou et al., 2007; Huang et al., 2004), have attracted more attention from researchers for recommending resources in folksonomy systems, in which only binary relationships such as user-resource relationships or user-tag relationships are considered (Yeung et al., 2008; Niwa et al., 2006; Tso-Sutter et al., 2008; Peng et al., 2010; Liang et al., 2008).
In reality, there exist tripartite relationships among users, resources and tags in folksonomies. Making full use of such relationships may improve the recall and precision of recommendation systems (Marinho et al., 2011). Let’s see a scenario: A, B and C are three users in CiteULike. A collects paper p1 and annotates it with two tags “text classification” and “feature selection”. B uses tag “text classification” to annotate paper p2. C collects the same paper as A, but C annotates this paper with two different tags “text mining” and “knowledge management” with A. In this case, users A and B are related to tag “text classification” and users A and C are connected by paper p1. Such case is very popular in folksonomies that different users are linked with the same resources or the same tags. It implies that the information on resources and tags is all useful for identifying the candidate resources when recommending, which inspires us to develop a model to clearly represent the tripartite relationships among users, resources and tags.
Before modeling, one main issue on data sparsity should be considered seriously, which is still a challenge for personalized recommendation. However, diffusion-based recommendation methods have demonstrated that they can effectively solve the data sparsity problem and meanwhile improve the accuracy and the diversification of recommendation systems (Zhou et al., 2007; Huang et al., 2004; Zhang et al., 2007). These kinds of methods suppose that user’s interest or taste can diffuse from itself to others. Zhang et al. (2010) developed a bipartite graph based taste diffusion model and combined two diffusion processes which conducted on user-resource bipartite graph and resource-tag bipartite graph respectively to recommend resources to users, but their research ignored the information between users and tags.