Stratified Sampling for Association Rules Mining

It is well recognized that mining association rules in a very large database is usually time consuming due to the I/O overhead in scanning the disk resident database. As one of the techniques for reducing the I/O overhead, sampling for mining association rules has been actively investigated during the last few years. Each sampling method and algorithm proposed in the literature has its own merits and demerits in terms of effectiveness and efficiency and none of them can claim to be the best. Which sampling method to use and how big the sample size should be for a given database are key issues in sampling for particular data mining tasks. In this paper a transaction size based stratified sampling method has been proposed, tested and compared with the simple random sampling method for mining association rules. It opens up the questions of how to stratify the datasets so that it can better suit the problem of association rule mining.