Statistical Supports for Frequent Itemsets on Data Streams

Abstract. A statistical technique is developed for estimating the support of itemsets on data streams, regardless of the size of the data stored. This technique, which is computationally ultra fast, does not depend on the algorithm used to build or maintain the itemsets. On frequent itemsets, it allows to maximize either the precision or the recall, as chosen by the user, while it does not damage the other criterion, and may even yield very good Fβ-measures. Since the maximization of both criteria is statistically hard, this provides algorithms building frequent itemsets with an efficient alternative to find those that are true frequents, when only a partial storing of the data stream is technically available. Experiments demonstrate the potential of the technique.

• Type: Conference
• Year: 2005
• Where: MLDM
• Authors: Pierre-Alain Laur, Jean-Emile Symphor, Richard Nock, Pascal Poncelet
• Keywords: Data Stream, Frequent Itemsets, Machine Learning, MLDM 2005, Statistical Technique