Information Filtering and Personalisation in Databases using Gaussian Curves G¨ unther Specht Technische Universit¨ at M ¨ unchen Institut f ¨ ur Informatik Orleansstr. 34, D-81667 M¨ unchen, Germany specht@in.tum.de Thomas Kahabka Exolution GmbH Barerstr. 60, D-80799 M¨ unchen, Germany thomas.kahabka@exolution.de Abstract We present an information filtering and adaptive person- alisation algorithm for arbitrary information systems based on databases. This algorithm is called GRAS (Gaussian Rating Adaptation Scheme), and it combines content-based and collaborative filtering. The goal is to filter retrieved documents of a query according to the personal interest of a user and to sort them according to the personal relevance. The algorithm tries to make the benefits of collaborative filtering available to application domains where collabo- rative filtering could not yet be applied due to lack of the critical mass of users or improper content structure. The algorithm collects background information about the user and the content by implicit and explicit feedback techniques. This information is then used to consecutively adapt user- and object profiles according their maturity. The described algorithm is applicable for the personalisation of any kind of application domain, even on multimedia data. GRAS is implemented in the multimedia database MultiMAP 1 as a generic personalisation provider module. 1 Introduction We are living in an age of information overflow. The vol- ume of information and the number of information sources are continuously increasing. In today’s world, it is al- most impossible to find and acquire relevant pieces of in- formation without being overflowed with irrelevant mate- rial. With personalisation methods we can improve the sit- uation. An important technique used in personalised systems is in- formation filtering which can be divided into three basic classes: content-based, collaborative and economic filter- ing. The wide research activities in information filtering 1 MultiMAP is funded by the “Deutsches Forschungsnetz Verein” (DFN) and the “Bundesministerium f ¨ ur Bildung, Wissenschaft, Forschung und Technologie” (BMBF) under contract TK 598-VA/V1. have resulted in several content-based and collaborative fil- tering systems. Most of them have been built for one spe- cial application and rely on a certain type of content. This prevents them from being used as a generic personalisation scheme in a hypermedia database. We needed to find a generic personalisation scheme for the multimedia database MultiMAP, which was developed in our research group at the Munich University of Tech- nology. This lead to the Gaussian Rating Adaptation Scheme (GRAS) which combines content-based and col- laborative filtering. Its efficient implementation using re- lational database technology allows online personalisation of hyperlinked multimedia objects. It is appropriate for the personalisation of multimedia and hyperlinked content, since it makes no assumptions about the structure of the ob- jects, i.e. objects do not have to be text. It is not dependent on a critical mass of users or objects to work effectively, as many collaborative filtering approaches do, which widens the area of use for GRAS. The rest of this paper is orga- nized as follows: In section 2 we give a short introduction to personalisation techniques. Section 3 presents the GRAS algorithm and in section 4 an efficient implementation us- ing relational database technology is discussed. Section 5 gives a brief overview of MultiMAP, the host hypermedia database system used to demonstrate GRAS. Finally in sec- tion 6 we discuss related work. 2 Personalisation in Information Systems The aim of personalisation is to select data whose con- tent are most relevant to the user from a greater volume of information and to present them in a suitable way for the user. The main logical parts (Fig 1) in a personalisation system are the user (and object) profile, user feedback and informa- tion filtering.