Graph Pattern Mining and Learning through User-defined Relations (Extended Version)

In this work we propose R-GPM, a parallel computing framework for graph pattern mining (GPM) through a user-defined subgraph relation. More specifically, we enable the computation of statistics of patterns through their subgraph classes, generalizing traditional GPM methods. R-GPM provides efficient estimators for these statistics by employing a MCMC sampling algorithm combined with several optimizations. We provide both theoretical guarantees and empirical evaluations of our estimators in application scenarios such as stochastic optimization of deep high-order graph neural network models and pattern (motif) counting. We also propose and evaluate optimizations that enable improvements of our estimators accuracy, while reducing their computational costs in up to 3-orders-of-magnitude. Finally,we show that R-GPM is scalable, providing near-linear speedups on 44 cores in all of our tests.