Significance measures and data dependency in classification methods

This paper is a comparative study to reveal the factors that have caused the contradictions among the previous comparative studies. The emphasis has been placed on the significance analysis of input variables, causality analysis and comparative biases. The role of attributes in the classification task with respect to their information bearing nature, and how this is approached by the classification algorithms are studied. The authors show that the ranked order of significance variables generally differs among the classification methods. The authors argue that data dependency is a key factor in the analysis of the contradictions among the previous comparative studies. The important role of class prototypes and significance measures, particularly the notion of positional causality introduced in this paper, facilitates future investigation on data transparency, particularly in neural networks. Positional causality means the ability to explain the significance of a variable in each partition of the problem space. Unlike the positional indicators, the authors argue that global indicators are not reliable. The authors show that global indicators do not satisfy the reasoning expectations of a human expert, but may cause confusion. The authors conclude that neural networks are reliable tools with which to analyse the positional causality of problem features.