Discovering Frequent Geometric Subgraphs

Abstract

Data mining-based analysis methods are increasingly being applied to datasets derived from science and engineering domains that model various physical phenomena and objects. In many of these datasets, a key requirement for their effective analysis is the ability to capture the relational and geometric characteristics of the underlying entities and objects. Geometric graphs, by modeling the various physical entities and their relationships with vertices and edges, provide a natural method to represent such datasets. In this paper we present gFSG, a computationally efficient algorithm for finding frequent patterns corresponding to geometric subgraphs in a large collection of geometric graphs. gFSG is able to discover geometric subgraphs that can be rotation, scaling, and translation invariant, and it can accommodate inherent errors on the coordinates of the vertices. We evaluated its performance using a large database of over 20,000 chemical structures, and our results show that it requires relatively little time, can accommodate low support values, and scales linearly with the number of transactions.