Hierarchical Pattern Discovery in Graphs

In this correspondence, an efficient technique for detecting repeating patterns in a graph is described. For this purpose, the searching capability of evolutionary programming is utilized for discovering patterns that are often repeating in such structural data. The approach adopted in this correspondence is hierarchical in nature. Once a pattern is discovered in a particular level of the hierarchy, the graph is compressed using it, and the substructure discovery algorithm is repeated with the compressed graph. The proposed technique is useful for mining knowledge from databases that can be conveniently represented as graphs. The importance of such an endeavor can hardly be overemphasized, given that substantial portion of data that are generated and collected is either structural in nature or is composed of parts and relations between the parts, which can be naturally represented as graphs. A typical example can be the structure of protein as well as computer-aided design circuits that have a natural graphical representation.

I. Introduction

Knowledge discovery [1]–[4] may be defined as the process of discovering interesting, potentially useful patterns from large datasets. Fig. 1 describes the overall process of knowledge discovery where data preparation, data mining, and knowledge representation are the three important tasks. In recent times, there has been a surge of activity aimed at the challenging task of discovering interesting patterns, concepts, and structural repetitions and making sense out of it in large amount of data that is generated and collected routinely. However, few pattern discovery techniques exploit the structural component in the data, either spatial on temporal, although much of the data collected is inherently structural in nature [5]. One method for discovering knowledge in the structural data is the identification of common patterns or concepts that describe interesting and repetitive substructures within the structural data. Once discovered, the substructure concept associated with the patterns can be used to simplify the data by replacing instances of the patterns with a pointer to the newly discovered concept. The discovered substructure concepts allow abstraction over detailed structure in the original data and provide new, relevant attributes for interpreting the data.