Automatic Assignment of Item Weights for Pattern Mining on Data Streams

Abstract

Research in Weighted Association Rule Mining (WARM) has largely concentrated on mining traditional static transactional datasets. Whilst there have been a few attempts at researching WARM in a data stream environment, none have addressed the problem of assigning and adapting weights in the presence of concept drift, which often occurs in a data stream environment. In this research we experiment with two methods of adapting weights; firstly, a simplistic method that recomputes the entire set of weights at fixed intervals, and secondly a method that relies on a distance function that assesses the extent of change in the stream and only updates those items that have had significant change in their patterns of interaction. We show that the latter method is able to maintain good accuracy whilst being several times faster than the former.