Interesting review, but I have to take exception to your last paragraph: I think Turchin is doing the right thing by only investigating a few selected variables (which he has substantial background reason for thinking of interest) as input into his models. Turning a neural network loose on every possible variable is just begging for massive datamining and multiple comparison problems which eliminate any validity you might hope to have for your results! Worse, if you use all your data initially, no one will be able to test your results for overfitting on any other data set...
Thanks for the feedback. I would guess you’re probably right. My knowledge of data mining practices is actually pretty minimal.
The review, however, was written for a class, and so it is academically mandatory (i.e. “If you want an A you better...”) to come up with problems with the original research and ways to improve. The professor seemed to like neural networks, so… (I think I inherited her “Just run everything through a neural network” mentality, but will definitely update my views based on your feedback. Thanks!)
I am interested in that review of yours.
It’s long, so I put it in dropbox. This link should take you there. (If not, let me know. My dropbox skills are probably sub-par)
Interesting review, but I have to take exception to your last paragraph: I think Turchin is doing the right thing by only investigating a few selected variables (which he has substantial background reason for thinking of interest) as input into his models. Turning a neural network loose on every possible variable is just begging for massive datamining and multiple comparison problems which eliminate any validity you might hope to have for your results! Worse, if you use all your data initially, no one will be able to test your results for overfitting on any other data set...
Thanks for the feedback. I would guess you’re probably right. My knowledge of data mining practices is actually pretty minimal.
The review, however, was written for a class, and so it is academically mandatory (i.e. “If you want an A you better...”) to come up with problems with the original research and ways to improve. The professor seemed to like neural networks, so… (I think I inherited her “Just run everything through a neural network” mentality, but will definitely update my views based on your feedback. Thanks!)