ML-DS: A Novel Deterministic Sampling Algorithm for Association Rules Mining

Due to the explosive growth of data in every aspect of our life, data mining algorithms often suffer from scalability issues. One effective way to tackle this problem is to employ sampling techniques. This paper introduces, ML-DS, a novel deterministic sampling algorithm for mining association rules in large datasets. Unlike most algorithms in the literature that use randomness in sampling, our algorithm is fully deterministic. The process of sampling proceeds in stages. The size of the sample data in any stage is half that of the previous stage. In any given stage, the data is partitioned into disjoint groups of equal size. Some distance measure is used to determine the importance of each group in identifying accurate association rules. The groups are then sorted based on this measure. Only the best 50% of the groups move to the next stage. We perform as many stages of sampling as needed to produce a sample of a desired target size. The resultant sample is then employed to identify association rules. Empirical results show that our approach outperforms simple randomized sampling in accuracy and is competitive in comparison with the state-of-the-art sampling algorithms in terms of both time and accuracy.