Salford Systems

CART Nominal (Ordered) Predictors and Rank

One of the strengths of CART is that, for ordered predictors, the only information CART uses are the rank orders of the data - not the actual values of the data. In other words, if you replace a predictor with its rank order, the CART tree will be unchanged.

This means that CART splits cannot be affected by any transform of the data that preserves order (monotone transform). For instance, AGE, log(AGE), AGE^2, etc., all would yield the same split. If you have a nominal variable with values 1,2,3,4,5: so long as the value "5" properly represents the highest value in the data and "1" represents the lowest value, and so forth, then monotone transforms of the data -- transforms that preserve rank ordering -- will not alter how the predictor acts in the tree.

Of course if a variable is categorical (discrete, unordered, nominal) then the values are just arbitrary labels. Simply indicating to CART that a predictor is variable is sufficient for proper handling.

November 2000.

All content above is copyright © 2000, Salford Systems. All rights reserved worldwide. The content of this site is made available to support the licensed use of Salford Systems data mining software products. No part of this site is to be copied, duplicated, modified or redistributed in whole or part without the prior written permission of Salford Systems.