|Authors:||Kriege, Nils Morten|
|Other Titles:||Algorithms & applications|
|Abstract:||Graphs are a well-studied mathematical concept, which has become ubiquitous to represent structured data in many application domains like computer vision, social network analysis or chem- and bioinformatics. The ever-increasing amount of data in these domains requires to efficiently organize and extract information from large graph data sets. In this context techniques for comparing graphs are fundamental, e.g., in order to obtain meaningful similarity measures between graphs. These are a prerequisite for the application of a variety of data mining algorithms to the domain of graphs. Hence, various approaches to graph comparison evolved and are wide-spread in practice. This thesis is dedicated to two different strategies for comparing graphs: maximum common subgraph problems and graph kernels. We study maximum common subgraph problems, which are based on classical graph-theoretical concepts for graph comparison and are NP-hard in the general case. We consider variants of the maximum common subgraph problem in restricted graph classes, which are highly relevant for applications in cheminformatics. We develop a polynomial-time algorithm, which allows to compute a maximum common subgraph under block and bridge preserving isomorphism in series-parallel graphs. This generalizes the problem of computing maximum common biconnected subgraphs in series-parallel graphs. We show that previous approaches to this problem, which are based on the separators represented by standard graph decompositions, fail. We introduce the concept of potential separators to overcome this issue and use them algorithmically to solve the problem in series-parallel graphs. We present algorithms with improved bounds on running time for the subclass of outerplanar graphs. Finally, we establish a sufficient condition for maximum common subgraph variants to allow derivation of graph distance metrics. This leads to polynomial-time computable graph distance metrics in restricted graph classes. This progress constitutes a step towards solving practically relevant maximum common subgraph problems in polynomial time. The second contribution of this thesis is to graph kernels, which have their origin in specific data mining algorithms. A key property of graph kernels is that they allow to consider a large (possibly infinite) number of features and can support graphs with arbitrary annotation, while being efficiently computable. The main contributions of this part of the thesis are (i) the development of novel graph kernels, which are especially designed for attributed graphs with arbitrary annotations and (ii) the systematic study of implicit and explicit mapping into a feature space for computation of graph kernels w.r.t. its impact on the running time and the ability to consider arbitrary annotations. We propose graph kernels based on bijections between subgraphs and walks of fixed length. In an experimental study we show that these approaches provide a viable alternative to known techniques, in particular for graphs with complex annotations.|
|Subject Headings:||Maximum common subgraph|
|Appears in Collections:||LS 11|
This item is protected by original copyright
All resources in the repository are protected by copyright.