Immune Inspired Information Filtering in a High Dimensional Space Nikolaos Nanas, Stefanos Kodovas, Manolis Vavalis, and Elias Houstis 1 Lab for Information Systems and Services Centre for Research and Technology - Thessaly (CE.RE.TE.TH) 2 Computing and Telecomunications Department, University of Thessaly {n.nanas,s.kodovas,m.vavalis,e.houstis}@cereteth.gr Abstract. Adaptive Information Filtering is a challenging computa- tional problem that requires a high dimensional feature space. However, theoretical issues arise when vector-based representations are adopted in such a space. In this paper, we use AIF as a test bed to provide ex- perimental evidence indicating that the learning abilities of vector-based Artificial Immune Systems are diminished in a high dimensional space. 1 Introduction The research domain of Adaptive Information Filtering (AIF), seeks to provide a solution to the problem of information overload, particularly on the Web, through the automatic construction of a representation of a user’s information interests, called “user profile”, and its continuous adaptation to temporal changes in these interests. The user profile is responsible for evaluating the relevance of new incoming information to the user’s interests and this assessment is exploited for providing the user with the appropriate information. More specifically, in content-based AIF, the user profile comprises descriptive features extracted from the content of relevant information items and these features are matched to those in new information items to assess their relevance. For example, in the case of textual information, which has been the main focus of research in AIF in general, and of the current work in particular, keywords extracted from the text of documents are used to abstract their content and to build the user profile. Typically, both the user profile and the documents are represented as binary, or weighted, keyword vectors in a space with as many dimensions as the number of unique keywords in the underlying document vocabulary. This allows the application of trigonometric measures of similarity for calculating how close to the profile’s vector is a document’s vector. Such vector-based representations has been the cornerstone of research in AIF, but have also been fundamental for research in the domain of Artificial Immune Systems (AIS). AIF is a complex and dynamic computational problem with no established solution. As we further discuss in [14], there is a characteristic lack of broadly adopted web applications that are based on content-based AIF, and this is in part due to its distinguishing and challenging requirements. One of these requirements