A Hybrid Interestingness Heuristic Approach for Attribute-Oriented Mining

Abstract

A hybrid interestingness heuristic algorithm, clusterAOI, is presented that generates a more interesting generalized final table than traditional attribute-oriented induction (AOI). AOI uses a global static threshold to generalize attributes irrespective of attribute features, consequently leading to overgeneralisation. In contrast, clusterAOI uses attribute features such as concept hierarchies and distinct domain attribute values to dynamically recalculate new attribute thresholds for each of the less significant attributes. ClusterAOI then applies new heuristic functions and the Kullback-leibler (K-L) measure to evaluate interestingness for each attribute and then for all attributes by a harmonic aggregation in each generalisation iteration. The dynamic threshold adjustment, aggregation and evaluation of interestingness within each generalization iteration ultimately generates a higher quality final table than traditional AOI. Results from real-world cancer and population datasets show both significantly increased interestingness and better performance compared with AOI.