Item

Abstract

Information filtering has been a research issue for years. In an information filtering scenario users information needs are expressed by user subscriptions, and users are notified about published documents or events that match these interests. The combination of the publish/subscribe scenario with the peer-to-peer (P2P) approach of autonomous peers makes high demands on the scalability and the efficiency of such a given highly distributed network. However, in many cases a subscriber is not interested in all the events that match his profile, but rather in a small representative set. In this paper, we present our approach of an approximate publish/subscribe system, that relaxes the assumption for receiving notifications from every information producer in the network. Our work builds upon distributed hash table technology to create and maintain a distributed global directory that contains information about peers' publishing behavior and combines the current peer state and the prediction of the future publishing behavior of a peer to store a subscription only to the most promising peers in the network. Our experimental evaluation shows that approximate information filtering results satisfying recall level and is able to accommodate changes in peer publishing behaviour.