CT-EXT: An algorithm for computing typical testor set

Abstract

Typical testors are a useful tool for feature selection and for determining feature relevance in supervised classification problems, especially when quantitative and qualitative features are mixed. Nowadays, computing all typical testors is a highly costly procedure; all described algorithms have exponential complexity. Existing algorithms are not acceptable methods owing to several problems (particularly run time) which are dependent on matrix size. Because of this, different approaches, such as sequential algorithms, parallel processing, genetic algorithms, heuristics and others have been developed. This paper describes a novel external type algorithm that improves the run time of all other reported algorithms. We analyze the behaviour of the algorithm in some experiments, whose results are presented here. © Springer-Verlag Berlin Heidelberg 2007.