Levelwise Search of Frequent Patterns with Counting Inference

In this paper, we address the problem of the eeciency of the main phase of most data mining applications: The frequent pattern extraction. This problem is mainly related to the number of operations required for counting pattern supports in the database, and we propose a new method, called pattern counting inference, that allows to perform as few support counts as possible. Using this method, the support of a pattern is determined without accessing the database whenever possible, using the supports of some of its sub-patterns called key patterns. This method was implemented in the Pascal algorithm that is an optimization of the simple and eecient Apriori algorithm. Experiments comparing Pascal to the Apriori, Close and Max-Miner algorithms, each one representative of a frequent patterns discovery strategy, show that Pascal improves the eeciency of the frequent pattern extraction from correlated data and that it does not induce additional execution times when data is weakly correlated.