Monday, September 03, 2007

"Data Mining" = voodoo science ?

On the Statistical Modeling blog, Aleks Jakulin has a rant on the virtues of data mining:

I view data analysis as summarization: use the machine to work with large quantities of data that would otherwise be hard to deal with by hand. I am also curious about what would the data suggest, and open to suggestions. Automated model selection can be used to list a few hypotheses that stick out of the crowd: I was not using model selection to select anything, but merely to be able to quantify how much a hypothesis sticks out from the morass of the null.

The response from several social scientists has been rather unappreciative along the following lines: "Where is your hypothesis? What you're doing isn't science! You're doing DATA MINING !"
I had almost the same reaction a while back when I was visiting JPL: the climatologists there were horrified at the idea of trolling for patterns in climate data, and to the person, asked me the dreaded 'But what is the science question?" question. Of course, given the general hot-potato-ness of climatology right now, one might sympathize with their skittishness.

Data mining is a tricky area to work in, and I've discussed this problem earlier. It's a veritable treasure-chest of rich algorithmic problems, especially in high dimensional geometry, and especially over large data sets. However, it's often difficult to get a sense of forward progress, especially since the underlying analysis questions often seem like elaborate fishing expeditions.

In that context, the distinction Aleks makes between confirmatory data analysis (check if the data validates or invalidates a hypothesis) and exploratory data analysis (play with the data to create a non-uniform distribution on plausible hypotheses) is quite helpful. It also emphasizes the interactive and very visual nature of data mining; interactive tools and visualizations are as important as the underlying analysis tools as well.

Update: Chris Wiggins points me to some of the earlier references to 'data mining'. One of the most vituperative is a paper by Michael Lovell in 1983 in The Review of Economics And Statistics. This paper drips with scorn for 'data miners', but makes a point that is at the very least worthy of consideration: namely that because of the large dimensionality of the space of hypotheses that a data mining application typically explores (here couched in terms of explanatory variables for a regression), patterns with apparently high p-values might not actually be that significant (or stated another way, in high dimensional spaces, there are many seemingly rare patterns that aren't that rare).

Disqus for The Geomblog