Knotenähnlichkeiten aus Aktivierungsausbreitungen

This thesis discusses how networks can be searched via methods of spreading activation. Networks, often represented as graphs, are datasets consisting of data objects (nodes or vertices) and the relations among them (edges). In many areas, e.g. information retrieval, networks are searched according to specific queries for related, interesting and relevant data objects, by applying methods such as spreading activation. In these methods, nodes, which represent the queries' data objects, are initially activated. The activation of each node subsequently spreads iteratively via connected edges to activate adjacent nodes with a certain level. The final level of activation after a certain number of iterations is interpreted as a heuristic measure of relevance, interestingness or importance. The result of the query is a list of data objects, sorted according to their level of activation. There are, however, several drawbacks to methods involving pure (constraint free) spreading activation, which is why until now, heuristic constraints have usually been applied to counteract these disadvantages. In this work another drawback, the convergence to a query-independent fix point is shown and discussed as well as methods to avoid this drawback without the usage of constraints. This work goes on to show how two kinds of similarities between nodes in graphs can be determined by applying the spreading activation process based on convergence. Sorting nodes according to both types of similarities to query nodes produces further possibilities of searching networks in addition to sorting based on levels of activation. Activation similarity is based on the overlap of the direct and indirect neighborhood of two nodes and is a relaxation of the maximal structural equivalence. The more overlap, the higher the degree of similarity of two nodes. Thus, two nodes with a high degree of similarity are reflected by a small distance in the graph, limiting the possibilities of searching based on activation similarity. This similarity measure is suitable for extracting dense surrounding subgraphs or communities surrounding query nodes. Signature similarity is based on comparing the structure of the direct and indirect neighborhoods of two nodes, and not on the overlap. It is shown that nodes that cannot be structurally distinguished, i.e. nodes that are automorphic images of each other, always have a maximal degree of similarity. The signature similarity is a relaxation of the maximal orbit equivalence and a heuristic measure of structural similarity of the neighborhood of nodes. Two nodes with a high degree of similarity …