News

I would like to offer the following thoughts in response to the letter from Arnold Goodman and your comments, regarding the ongoing debate over data mining and statistics.

The key to successful data mining is domain understanding, not statistics. Remember the lessons from KDD Cup 2000? Data mining packages can discover, and statistical analysis can confirm, all kinds of fabulous connections that we think will make powerful predictions. A little domain knowledge prompts us to ask the right questions which "peel back the onion," leading us to the true relationship, and we find that our exciting discovery loses its luster. The ability to ask the right questions comes not from a deep understanding of statistics or machine learning algorithms, but from a broad-based knowledge of the problem, the data, and the mechanisms by which they interact. In my field it is crucial to understand the mortgage process, the industry structure, the industry history, and borrower psychology to be able to generate useful analysis and models. Statistics are a necessary part of developing that knowledge, but they are not sufficient.

Data mining practitioners are, in my experience, extremely pragmatic about their work. In contrast, my experience with statisticians is that they are often too concerned with proofs and tidiness. In a world where we are measured by quarterly impact on the bottom line, working models are preferred to statements about the statistical significance of a coefficient. In the end, the pragmatism that has driven data mining will ensure its longevity, and it will ensure that useful methods and principles, whether derived from machine learning or from statistics, will endure, too. The best data miners will have a functional command of those methods and the ability to quickly acquire significant domain expertise.

Ted Durant Manager, Analytic Services Mortgage Guaranty Insurance Corp. Milwaukee, WI USA

Disclaimer: The opinions expressed in this document are the views of the author and are in no way intended to represent the views of his employer.