Change itemset mining in data streams

Abstract

Data stream mining is becoming very important in many application areas, such as the stock market. A data stream consists of an ordered sequence of instances and because there are usually a large number of instances along with limited computing and storage capabilities, algorithms that read the data only once are preferred. There has been some research that focuses on finding when a concept has changed, given some knowledge about the previous instances in the data stream, but little on determining the characteristics of that change. In this paper we concentrate on finding the characteristics of the changes that occur, using frequent itemset mining techniques. We propose an approach, which combines both heuristic and statistical approaches, that analyses changes that have occurred within a stream at itemset level and identify three types of change: extension, reduction, and support fluctuation. We evaluate our algorithm using both synthetic and real world datasets. © Springer-Verlag 2013.