High-performance parallel frequent subgraph discovery

Discovery of frequent subgraphs of an input network is one of the most important facilities for mining and analyzing complex networks. The most accurate solution to frequent subgraph discovery is to enumerate all subgraphs of size k and then count the frequency of each isomorphic class. However, the process is much time consuming because the number of subgraphs grows exponentially with the growth of the input network, or by increasing the size of the subgraphs. Also, there is no known polynomial-time algorithm for subgraph isomorphism detection, and this issue makes the problem harder. Hence, the available solutions can just mine small input networks and small subgraph sizes. A parallel and load-balanced solution named Subdigger is proposed which is faster and more efficient compared to available solutions. Subdigger efficiently executes on current multicore and multiprocessor machines, and incorporates a fast heuristic with a high-performance concurrent data structure which significantly accelerates detection and counting of isomorphic subgraphs. Subdigger can also handle large networks and subgraph sizes using external memory and external sorting. We performed several experiments using real-world input networks. Compared to the available solutions, Subdigger can extract frequent subgraphs much faster and the performance scales almost linearly using additional processor cores. The experimental results show that Subdigger can be more than 100 times faster than other solutions on a 4-core Intel i7 machine. Besides performance, Subdigger can process larger subgraphs using external memory while other tools crash due to memory limitation.