BAR: bitmap-based association rule: an implementation and its optimizations

The association rule mining, one of the most popular data mining techniques, is to find the frequent itemsets which occur commonly in transaction database. Of the various association algorithms, the Apriori is the most popular one, and its implementation technique to improve the performance has been continuously developed during the past decade. In this paper, we propose a bitmap-based association rule technique, called BAR, in order to drastically improve the performance of the Apriori algorithm. Compared to the latest Apriori implementation, our approach can improve the performance by nearly up to two orders of magnitude. This gain comes mainly from the following characteristics of BAR: 1) bitmap based implementation paradigm, 2) reduction of redundant bitmap-AND operations, and 3) an efficient implementation of bitmap-AND and bit-counting operation by exploiting the advanced CPU technology, including SIMD and SW prefetching. We will describe the basic concept of BAR approach and its optimization techniques, and will show, through experimental results, how each of the above characteristics of BAR can contribute the performance improvement.