Implementing Relevance Feedback as Convolutions of Local Neighborhoods on Self-Organizing Maps Markus Koskela, Jorma Laaksonen, and Erkki Oja Laboratory of Computer and Information Science, Helsinki University of Technology P.O.BOX 5400, 02015 HUT, Finland {markus.koskela,jorma.laaksonen,erkki.oja}@hut.fi Abstract. The Self-Organizing Map (SOM) can be used in implement- ing relevance feedback in an information retrieval system. In our ap- proach, the map surface is convolved with a window function in order to spread the responses given by a human user for the seen data items. In this paper, a number of window functions with different sizes are compared in spreading positive and negative relevance information on the SOM surfaces in an image retrieval application. In addition, a novel method for incorporating location-dependent information on the relative distances of the map units in the window function is presented. 1 Introduction The data organization provided by the Self-Organizing Map (SOM) [1] can be utilized in searching for interesting data items. Due to the topology-preservation property of the SOM, neighboring map units contain similar feature vectors. If we already know that certain map units contain data items which are in some manner similar to the item we are interested in, a natural strategy is to focus the further search in the neighborhoods of these map units. This kind of setting arises, e.g. in iterative multi-round information retrieval where, on each query round, the user marks the retrieved items as relevant or nonrelevant to the query. The system then uses this information in estimating what the user is looking for. This kind of iterative refinement of a query is known as relevance feedback in information retrieval literature [2]. Content-based image retrieval (CBIR) has been a subject of recent intensive research effort. It differs considerably from textual information retrieval as, unlike text that consists of words, images do not consist of such basic building blocks which could directly be utilized in retrieval applications. Instead, the retrieval is based on visual features extracted from the images and alternative retrieval paradigms must be used. One common approach is query by example, where the user specifies her object of interest by giving or pointing out examples of interesting or relevant images. Relevance feedback is essential here, as the systems are normally not capable of returning the desired image on the first query round [3]. A CBIR system implementing J.R. Dorronsoro (Ed.): ICANN 2002, LNCS 2415, pp. 981–986, 2002. c  Springer-Verlag Berlin Heidelberg 2002 © 2002 Springer-Verlag. Reprinted with permission from Proceedings of the International Conference on Artificial Neural Networks (ICANN 2002). Madrid, Spain, pages 981-986.