Comparing two discriminant probabilistic interestingness measures for association rules

Abstract

Preliminary nomalization is needed for probabilistic pairwise comparison between attributes in Data Mining. Normalization plays a very important part in preserving the discriminant property of the probability scale when the number of observations becomes large. Asymmetrical associations between boolean attributes are considered in our paper. Its goal consists of comparison between two approaches. The first one is due to a normalized version of the "Likelihood Linkage Analysis" methodology. The second one is based on the notion of "Test Value" defined with respect to a hypothetical sample, sized 100 and summarizing the initial observed sample. Two facets are developed in our work: theoretical and experimental. A comparative experimental analysis is presented with the well known databases "Wages" and "Abalone". © Springer-Verlag Berlin Heidelberg 2013.