Recommending WWW information sources using Feature Guided Automated Collaborative Filtering Gabriela Polčicová and Pavol Návrat Slovak University of Technology, Department of Computer Science and Engineering, Ilkovičova 3, 812 19 Bratislava, Slovakia {polcicova, navrat}@dcs.elf.stuba.sk Abstract. Exponential growth of the number of in- formation sources available on the web hampers ef- fective identification of those items that are likely to be of interest to the user. To overcome the problem of information overload, techniques of information fil- tering aim at recommending items of potential interest based on a profile of user needs, preferences or opin- ions. In the paper, a recommender system is described that utilizes Feature Guided Automated Collaborative Filtering for recommending relevant HTML- documents to the users. While browsing the web, user expresses his/her opinions on documents by rating them. The system "learns" user's opinions and searches for like-minded users in order to recommend him unseen relevant documents of interest. The sys- tem’s architecture is based on two kinds of agents - a recommendation agent and a communication agent - that work for each user. 1 Introduction Exponential growth of the number of information sources available on the web hampers effective identi- fication of those items that are likely to be of interest to the user. To overcome the problem of information overload, techniques of information filtering aim at recommending items of potential interest based on a profile of user needs, preferences or opinions on one hand, and based on an analysis of the item on the other hand. Recommander systems can either analyse informa- tion items by themselves, employing one of the tech- niques of content-based filtering, or rely on opinions of other users and employing one of the collaborative filtering techniques. Techniques for recommending appropriate information using opinions of users with similar preferences are known as Automated Collabo- rative Filtering (ACF) [7, 9]. The recommender sys- tem that is described in this paper uses a particular kind of ACF called Feature Guided Automated Col- laborative Filtering (FGACF). We shall shortly describe the technique in the next section. In section 3, we shall present the recom- mender system that we currently work on. It should be noted that our work makes use of some results of G. Polčicová’s masters thesis that have been presented in [6]. 2 Feature Guided Automated Collaborative Filtering ACF approach is based on the hypothesis that users who had similar preferences in the past will probably have similar preferences again [8]. Recommender systems that use ACF collect users’ opinions on a set of items in the form of their ratings. They search through the community of users to identify users with similar opinions. Similarity is defined in terms of ratings of respective items. Opinions of those like- minded users about items that a particular user with a similar interest has not seen are used to predict his/her ratings. Items with positively predicted ratings are recommended to the user. An important advantage of this approach, as com- pared to other filtering techniques, is that content of items being filtered need not be analysed at all. With regard to the assumption that agreement of prefer- ences in one topic (e.g. in literature) does not imply agreement of preferences in another topic (e.g. in sports), it is necessary to recommend items of each topic (category) separately. This method of recom- mendation is known as Feature Guided Automated Collaborative Filtering (FGACF). A variety of algorithms and systems for ACF have been reported in the literature. Most of the systems, such as the system Tapestry [4] that filters e-mails, Ringo [9] recommending music albums and artists, GroupLens [7] using Usenet newsgroups as a domain,