Using count prediction techniques for mining frequent patterns in transactional data streams

We study the problem of mining frequent itemsets in dynamic data streams and consider the issue of concept drift. A count-prediction based algorithm is proposed, which estimates the counts of itemsets by predictive models to find frequent itemsets out. The predictive models are constructed based on the data in the data stream and serve as a description of the concept of the stream. If there is a concept drift in the stream, the description of the concept can be updated by reconstructing the predictive models. According to our experimental results, the proposed algorithm is efficient and has stable performance. Besides, using respective predictive models for count-predictive mining would preserve the quality of mining answers effectively (in terms of accuracy) against the change of the concept.