May

31

When you use many different cutoff points for many different numbers, giving yourself the equivalent of 100s of splits of the data, you are guaranteed by randomness to find a scenario that is at one extreme or the other. This is the reason that what they used to call the automatic interaction detector, (which has another name now ) is totally useless as a predictor. Maybe it should be called the artificial interaction detector.

anonymous shares: 

The Automatic Interaction Detector (AID) (and its ilk CHAID, CART, etc.) is apparently still frequently used in the analysis of marketing data (for example to discover what characteristics your buyers have in common).

It goes back to the 1960s and 1970s (Sonquist, Baker & Morgan), and some of its weaknesses were realized even then:

Doyle: "THE USE OF AUTOMATIC INTERACTION DETECTOR AND SIMILAR SEARCH PROCEDURES"
 


Comments

Name

Email

Website

Speak your mind

Archives

Resources & Links

Search