Direct out-of-memory distributed parallel frequent pattern mining

Frequent itemset mining is a well studied and important problem in the datamining community. An abundance of different mining algorithms exists, all with different flavor and characteristics, but almost all suffer from two major shortcomings. First, in general frequent itemset mining algorithms perform exhaustive search over a huge pattern space. Second, most algorithms assume that the input data fits into main memory. The first problem was recently tackled in the work of [2], by direct sampling the required number of patterns over the pattern space. This paper extends the direct sampling approach by casting the algorithm into the MapReduce framework, effectively ceasing the memory requirements that the data should fit into main memory. The results show that the algorithm scales well for large data sets, while the memory requirements are solely dependent on the required number of patterns in the output.