| The focus of this thesis is to apply relational database techniques to accommodate all aspects of graph mining. Our primary goal is to address scalability of graph mining to very large data sets, riot currently addressed by main memory approaches. This thesis addressed the most general graph representation including multiple edges between any two vertices, and cycles. This thesis extends previous work (EDB-subdue) in a number of ways: improved substructure representation to avoid false positives during frequency counting, unconstrained substructure expansion with pseudo duplicate elimination for expanding multiple edges, canonical ordering of substructures for getting true count, hierarchical reduction for producing abstract pattern and generalization of DMDL that includes the presence of multiple edges in a subgraph. We also extend the substructure pruning to include ties when selecting top beam substructures. (Abstract shortened by UMI.)... |
