I thought i'd start up a discussion thread .
(Assuming anyone else on this little geek's forum is a fan or works with or a researcher in anything involving machine learning, data science, data mining, or touches on it )
So, how about them Random forests huh? And did you see the size of that multilayer convolutional net? And can you believe the deep learning perturbation on that backprop!
Photography Genius is always allowed some leeway, once the hammer has been pried from its hands and the blood has been cleaned up.
To improve is to change; to be perfect is to change often.
In 11 days my semester begins, and one of the classes I am taking is on data mining/machine learning. The focus of the class is mostly on the applications of the methods to genomics research.
I'm just about done on a smart options pricing (low-latency exchange traded) project and will be gearing up for a cross-asset deep risk analysis project next. It will be moderately large data (millions of live and billions of historic transactions), haven't got into the analysis yet so don't know what techniques I will be deploying. Probably not machine learning but definitely tweaking/upgrading quant models for more modelling power, e.g. at minimum basic stuff like implementing stratified monte carlo instead of naive monte carlo. Still working on a narrative extraction project in my spare time, but that is ultra-unfashionable-unhip-untrendy-unmentionable probablistic-symbolic not-big-data-at-all.
Is there any reason to think of a random forest technique as anything other than just a special case of a bootstrap/Monte Carlo method? Just briefly skimming naively through a description of the method it just seems like an application of the bootstrap to decision tree data rather than a standalone method in its own right.
I wouldn't really know enough to be able to comment.
All my random-forest mucking around was trying to predict binary data given a bunch of input variables. I did muck around with weighting the votes of each tree in the forest trying to minimise RMS error of the weighted vote across the training set (false being zero, true being 1), and counting instances that were in-bag for a given forest as 0.5. It did seem to give a more accurate forest (lower false positive and false negative rates) than equally weighting the vote.
A mad person thinks there's a gateway to hell in his basement. A mad genius builds one and turns it on. - CaptainChewbacca
. 
