Dependance of critical dimension on learning machines and ranking methods
Divya Suryakumar, Andrew H. Sung, Qingzhong Liu · 2012
Feature reduction is a major problem in data mining. Though traditional methods such as feature ranking and subset selection have been widely used, there has been little consideration given to assuring satisfactory performance of a learning machine in relation to the minimum of features required or the “critical dimension”. This critical dimension is unique to a specific dataset, learning machine, and ranking algorithm combination. The empirical methods demonstrate that many datasets show the existence of critical dimension. The dependence of this critical dimension on the learning machines and ranking algorithms could provide newer insights in understanding datasets, machine learning classifiers and ranking algorithms. The preliminary results of analysis show that the critical dimension depends largely on the feature ranking algorithm and that learning machines are less significant in determining the critical dimension.