A Statistical Method for Determining Importance of Variables in an Information System

摘要

A new method for estimation of attributes' importance for supervised classification, based on the random forest approach, is presented. Essentially, an iterative scheme is applied, with each step consisting of several runs of the random forest program. Each run is performed on a suitably modified data set: values of each attribute found unimportant at earlier steps are randomly permuted between objects. At each step, apparent importance of an attribute is calculated and the attribute is declared unimportant if its importance is not uniformly better than that of the attributes earlier found unimportant. The procedure is repeated until only attributes scoring better than the randomized ones are retained. Statistical significance of the results so obtained is verified. This method has been applied to 12 data sets of biological origin. The method was shown to be more reliable than that based on standard application of a random forest to assess attributes' importance.