An optimal attribute quantization algorithm for mining quantitative association rules

Abstract

Mining quantitative association rules on numerical attributes requires to partition quantities of each numerical attribute into a set of disjoint intervals, interpret each interval as an item, and then apply standard boolean association rule mining .In this paper we propose an optimal algorithm for partitioning numerical attribute values that runs in O (m )sequential time and O (log m )time using O (m/log m )EREW processors respectively for m attribute values. Without loss of generality that our linear-scan merge works on any distance metric, we use the maximal intra-interval distance metric rather than average distance, taking into consideration of the special properties of association rule coverage. Our algorithm partitions attribute values into intervals with maximal intra-interval distance within a given threshold and keeps the populations among intervals as equal as possible.