Neighbor Selection and Recommendations in Social Bookmarking Tools Antonina Dattolo, Felice Ferrara and Carlo Tasso Dipartimento di Matematica ed Informatica Universit` a degli Studi di Udine Via delle Scienze, Udine, Italy {antonina.dattolo,felice.ferrara,carlo.tasso}@dimi.uniud.it Abstract Web 2.0 applications innovate traditional informative services providing Web users with a set of tools for pub- lishing and sharing information. Social bookmarking sys- tems are an interesting example of this trend where users generate new contents. Unfortunately, the growing amount of available resources makes hard the task of accessing to relevant information in these environments. Recommender systems face this problem filtering relevant resources con- nected to users’ interests and preferences. In particular, collaborative filtering recommender systems produce sug- gestions using opinions of similar users, the neighbors. The task of finding the neighbors is difficult in environment such as social bookmarking systems, since bookmarked re- sources belong to variegated domains. In this paper we propose a methodology for partitioning users, tags and resources into domains of interest. Filtering tags and resources in accordance to the specific domains we can select a different set of neighbors for each domain, improving the accuracy of recommendations. 1. Introduction Web 2.0 applications innovate traditional informative services providing Web users with a set of tools for pub- lishing and sharing information. This process shifts the task of generating contents from a selected and restricted set of authors to a new, wide population of publishers. An interesting example of this trend, which is going to in- novate production and access to information, is represented from social bookmarking systems. These systems allow users to collect resources assigning them a set of tags. Often users employ the tags as indices for re-finding the resources which they have previously visited. The other users could also benefit from this process; in fact, tags offer a personal, meta description of the resources and can be used for find- ing peers with similar interests. Unfortunately, the numeric explosion of generated re- sources makes this task difficult. Search engines, mainly Google TM , are the most used tools for finding document on the Web, but they do not return personalized results; in par- ticular they do not take in account users’ preferences and goals. Recommender systems [1] filter resources using informa- tion stored in a user model. In particular, collaborative filtering recommender systems compute similarity among users, identifying for each user the set of her/his neighbors (i.e. peers with similar preferences and features), and then suggest her/him new resources considering the set of re- sources visited by neighbors. So, collaborative filtering recommender systems apply the following two steps: 1. Neighbor selection. The behavior of a user A is com- pared with the behavior of other users in order to find a set of neighbors. 2. Word-of-mouth simulation. The resources, identi- fied as interesting for neighbors of A, are suggested to her/him, simulating a word-of-mouth process. In order to execute the process of neighbor selection, users’ behaviors have to be modeled. The user modeling process is based on the acquisition of users’ ratings, where each rat- ing is an association between a user and an item described by means of some value [14]. Some applications, such as Movielens 1 , allow users to evaluate a resource in an explicit way, for example, by means of the assignation of a vote. Other environments, such as social bookmarking systems, can infer users’ ratings implicitly, considering bookmarks as a manifestation of interest. Focusing our attention on social bookmarking systems the neighbor selection can be implemented evaluating the num- ber of resources shared among users. However, collabo- rative filtering recommender systems work well when the resources belong to a same domain, since the process of 1 http://movielens.umn.edu/