Mining Graph Pattern Association Rules

Abstract

We propose a general class of graph-pattern association rules ($\mathsf {GPARs}$) for social network analysis, e.g., discovering underlying relationships among entities in social networks. Despite the benefits, $\mathsf {GPARs}$ bring us challenges: conventional support and confidence metrics no longer work for $\mathsf {GPARs}$, and discovering $\mathsf {GPARs}$ is intractable. Nonetheless, we show that it is still feasible to discover $\mathsf {GPARs}$. We first propose a metric that preserves anti-monotonic property as support metric for $\mathsf {GPARs}$. We then formalize the $\mathsf {GPARs}$ mining problem, and decompose it into two subproblems: frequent pattern mining and $\mathsf {GPARs}$ generation. To tackle the issues, we first develop a parallel algorithm to construct DFS code graphs, whose nodes correspond to frequent patterns. We next provide an efficient algorithm to generate $\mathsf {GPARs}$ by using DFS code graphs. Using real-life and synthetic graphs, we experimentally verify the performance of the algorithms.