Spatial Organization Using Self-Organizing Neural Networks Riccardo Rizzo, Marco Arrigo Italian National Research Council-Institute for Educational and Training Technologies Via Ugo La Malfa 153, 90146 Palermo,Italy {rizzo,arrigo}@itdf.pa.cnr.it ABSTRACT Spatial hypertext systems use physical properties as color, dimensions, and position to represent relationships between documents. These systems allows the user to express a lot of different relationships between information but the structure should be build by hand by the user. This can be complex if a large number of information is involved. Self-organizing neural networks map can automatically generate a document map in which clusters of similar documents are organized. These maps can be used as a navigation tool “per se” or as a starting point for more complex spatial organizations. Systems based on SOM network can also automatically find the right map location for a new document, giving to the user a valuable help in information organization. In this paper the application of Self-Organizing Maps as a tool to develop information maps and spatial hypertext systems prototype is discussed and some applications are presented. Keywords Self-organizing networks, neural networks, visualization. 1. INTRODUCTION In spatial hypertexts the user creates a two-dimensional space where documents and pieces of information are organized, and their relationships are expressed by using their relative location on a two dimensional space [5]. The freedom in building this information space allows the user to express a lot of different relationships using a “constructive ambiguity” [8] that is one of the major advantages of spatial hypertexts. In this paper spatial hypertexts refers to a set of systems that uses a spatial organization to represent a semantic structure. An attempt to integrate statistical representation of the free text of a document and spatial organization was made in [2]. This approach brings to a system that is difficult to scale and to a visualization of the information space difficult to manage. In this paper we present another approach to automatic organization of information; this approach is based on a self-organizing neural network that automatically build a useful information map. Self-Organizing Maps (SOM maps) [3] are artificial neural networks that can organize information or documents on a space using a two-dimensional array of neurons. The neurons can be considered as sensitive units capable to modify a set of parameters (their weights) in order to approximate an external input, during the learning stage. The mechanism of the learning stage will be explained in the following sections but in an informal way we can say that the when a document (represented by a set of suitable parameters) is submitted to the network the unit most sensitive (most similar to the document) usually called b.m.u. (best matching unit) is modified in order to become “more similar” (i.e. more sensitive) to the document itself. Different units can adapt their weights to different documents and on the surface of the map many specialized areas (set of units) will appear. This mechanism can be used for visualization and for organization of a set of document exploiting the same visual mechanism that is exploited in spatial hypertexts. In the paper the application of SOM map to visual information organization and visualization is described. The results obtained show that the SOM is a viable help to automatically organize information on a two-dimensional space. In the next section the principle of the Self-Organized Map is explained, then the application prototypes are presented and some conclusions are drawn. 2. SELF-ORGANIZING NEURAL NETWORKS TO DEVELOP INFORMATION MAPS Artificial neural networks (ANN) models are made up of a dense interconnection of simple non-linear computational elements corresponding to the biological neurons. Each connection is characterized by a variable weight that is adjusted, together with other parameters of the net, during the so-called "learning stage". In Self-Organized Map the neurons are organized in a lattice, usually a one or two-dimensional array, which is placed in the input space and is spanned over the inputs distribution. Using a two-dimensional SOM network it is possible to obtain a map of input space where closeness between units or clusters in the map represents closeness of the input vectors. The SOM algorithm principle can be explained in an abstract system without reference to any biological structure. To each processing unit in the SOM lattice is associated a vector of weight of the same dimension of the input vectors. Using the weights of each processing unit as a set of coordinates the lattice can be positioned in the input space. During the learning stage the weights of the units change their position and "move" towards the input points as illustrated in figure 1. This "movement" becomes slower and at the end of the learning stage the network is "frozen" in the input space.