抄録
Recently, pattern mining in structured domain, such as sequences, trees and graphs, is becoming increasingly abundant and several algorithms for especially frequent pattern mining have been developed. On the other hand, the research area of correlation mining in transaction databases, that extracts the underlying dependency among objects, attracts a big attention and extensive studies have been reported. Although we can easily expect to get a more powerful tool for structured data by introducing correlation mining, the most of current research on correlation mining are designed for transaction databases and little attention is paid to mining correlations from structured data. Motivated by these backgrounds, in this paper, we bring the concept of hyperclique pattern in transaction databases into the graph mining and consider the discovery of sets of highly-correlated subgraphs in graph-structured databases. To achieve this objective, a novel algorithm named HSG is proposed. By considering the generality ordering on sets of subgraphs, HSG employs the depth-first/breadth-first search strategy with powerful pruning techniques based on both of the anti-monotone property of support value and the upper bound of h-confidence measure. Experiments with artificial and real world datasets were conducted to assess the effectiveness of the proposed algorithm. The results of experiments show that HSG succeeds in discovering sets of highly-correlated subgraphs within reasonable computation time.
