Mining Frequent Itemsets from Secondary Memory

Mining frequent itemsets is at the core of mining association rules, and is by now quite well understood algorithmically. However, most algorithms for mining frequent itemsets assume that the main memory is large enough for the data structures used in the mining, and very few efficient algorithms deal with the case when the database is very large or the minimum support is very low. Mining frequent itemsets from a very large database poses new challenges, as astronomical amounts of raw data is ubiquitously being recorded in commerce, science and government. In this paper, we discuss approaches to mining frequent itemsets when data structures are too large to fit in main memory. Several divide-and-conquer algorithms are given for mining from disks. Many novel techniques are introduced. Experimental results show that the techniques reduce the required disk accesses by orders of magnitude, and enable truly scalable data mining.