Monday, February 25, 2013

Data analysis, interpretation and explanations

There was a recent data-related kerfuffle between the New York Times and the makers of the Tesla electric car. If you haven't read the articles, (and the NYT public editor's post on this has good links), the crude summary is this:

  • NYT reporter takes Tesla on long test drive, and reports problems with running out of charge.
  • Tesla Motors CEO Elon Musk writes a long takedown of the reporter's review, complete with graphs generated from data the car recorded during the drive
  • NYT comes back with their own interpretation of data, rebutting Musk's claims.
  • Others attempt to reproduce the reporter's experience and fail, but arguably in different weather conditions that might or might not make a difference.
In an insightful meta-analysis of the dustup, Taylor Owen of the Tow Center for Digital Journalism discusses the implications for journalism in a data-driven world. He also references an article by David Brooks that makes the point:
People are really good at telling stories that weave together multiple causes and multiple contexts. Data analysis is pretty bad at narrative and emergent thinking...
I've felt for a while now that it's time to design mechanisms for providing context and narrative to data analysis*. Some of the research my student Parasaran does is on metaclustering: essentially the meta-analysis of different perspectives (clusterings) of data to draw out a larger meaning. We've also just submitted a paper on how to derive local markers of prediction validity in clustering (rather than just relying on global measures of quality). And my seminar this semester is about the larger problems of explanations and accounting in data mining.

I think as computer scientists, we have a lot to offer in the realm of data mining - not just in the design of tools for prediction, but in the design of tools to facilitate better understanding.


* None of this is surprising to experimental scientists, who will routinely attack a problem from multiple directions in order to acquire a larger understanding of a phenomenon rather than just the ability to predict. 


Post a Comment

Disqus for The Geomblog