Information filtering via Iterative Refinement

– With the explosive growth of accessible information, expecially on the Internet, evaluation-based filtering has become a crucial task. Various systems have been devised aiming to sort through large volumes of information and select what is likely to be more relevant. In this letter we analyse a new ranking method, where the reputation of information providers is determined self-consistently. Introduction. – The study of complex networks and of some dynamical processes taking place on these structures has recently attracted a great deal of attention in the physics community [1–4]. The importance of technological networks, such as the Internet, lies mostly in the increased communication capabilities [5, 6], which make information progressively easier to produce and distribute. As storage and transmission costs continue to drop, an overabundance of information threatens to overwhelm its recipients. It is, therefore, crucial to process information in order to present a user only the one that answers best her requests [7]. An important aspect of information filtering regards scoring systems in the World Wide Web [8,9]. They collect evaluations and aggregate them into published scores that are meaningful to the final user. This embraces many different instances, ranging from commercial websites, where buyers evaluate sellers (Ebay, Amazon, etc.) to new generation search engines (Google, Yahoo, etc.), and opinion websites, where people evaluate objects (Epinions, Tailrank, etc.) Since the evaluators carry different expertise, it is important to estimate how accurate a given vote may be and to weight it accordingly. This can be done through the use of raters’ reputations [10]. Reputation summarises one’s past behaviour and has always been used to bear the risk of interacting with strangers. The Internet, while enhancing such a risk, brings in the possibility to find its antidotes [11]. Since nobody knows a-priori who are the honest and competent evaluators, in fact, online scoring systems often include some measure of their past performance. This gives users an indication on how trustworthy a given piece of information is supposed to be. An expert of the field would probably obtain a high reputation; experts’ votes should then count more when aggregating the scores. While reputation is usually obtained by asking users supplementary evaluations about other users, the procedure of Iterative Refinement (IR), which can be shown to outperform naive methods [12], does not require to explicitly rate the raters.