Clustering transactions with an unbalanced hierarchical product structure

Abstract

The datasets extracted from large retail stores often contain sparse information composed of a huge number of items and transactions, with each transaction only containing a few items. These data render basket analysis with extremely low item support, customer clustering with large intra cluster distance and transaction classifications having huge classification trees. Although a similarity measure represented by counting the depth of the least common ancestor normalized by the depth of the concept tree lifts the limitation of binary equality, it produces counter intuitive results when the concept hierarchy is unbalanced since two items in deeper subtrees are very likely to have a higher similarity than two items in shallower subtrees. The research proposes to calculate the distance between two items by counting the edge traversal needed to link them in order to solve the issues. The method is straight forward yet achieves better performance with retail store data when concept hierarchy is unbalanced. © Springer-Verlag Berlin Heidelberg 2007.