Efficiently extracting frequent subgraphs using MapReduce

Frequent subgraph extraction from a large number of small graphs is a primitive operation for many data mining applications. To extract frequent subgraphs, existing techniques need to enumerate a large number of subgraphs which is superlinear with the cardinality of the dataset. Given the rapid growing volume of graph data, it is difficult to perform the frequent subgraph extraction on a centralized machine efficiently. In this paper, we investigate how to efficiently perform this extraction over very large datasets using MapReduce. Parallelizing existing techniques directly using MapReduce does not yield good performance as it is difficult to balance the workload among the compute nodes. We therefore propose a framework that adopts the breadth first search strategy to iteratively extract frequent subgraphs, i.e., all frequent size-(i+1) subgraphs are generated based on frequent size-i subgraphs at the i^th iteration using a single MapReduce job. To efficiently extract frequent subgraphs, we propose an isomorphism-testing-free approach by properly maintaining how frequent subgraphs are mapped within each graph. Extensive experiments conducted on our in-house clusters demonstrate the superiority of our proposed solution in comparison with the baseline approach.