An assertive will for seeing and believing: introducing a feature cardinality driven distance measure to uninformative distributions
Joan Garriga
In: QIMIE'09, 27-30 April 2009, Bangkok, Thailand.

Abstract

What regard should a learning algorithm hold for the different information traces found in a sample? Answering this question objectively is not easy. Moreover, given that a full range of traits can be found in a human learning analogy, from the most daring or ingenuous, to the most conservative or incredulous. But in AI domains it is a must to clearly state the right will for believing what is seen when mining data bases. A key concept in this matter is assertiveness. The aim of this work is to ponder an approach to assertive KDDB, based on a feature cardinality driven distance measure to uninformative distributions. From this perspective, we present an alternative option to the support-confidence framework. The biases of this measure have not yet been thoroughly studied but the measure itself has proved to be quite effective as a heuristic when searching to optimize a sample in a simultaneous multi-interval discretization of continuous features. The empirical results show that the most relevant association or classification rules are revealed. Also, optimal cardinalities and optimal subsets of parents are found for any feature, according to a natural bias toward the MDL principle. As a conclusion, it appears the measure assertively captures knowledge. This may be useful for other data mining issues.

PDF - PASCAL Members only - Requires Adobe Acrobat Reader or other PDF viewer.

[Edit]