Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: Kernel methods / machine learning
From
Nick Cox <[email protected]>
To
[email protected]
Subject
Re: st: Kernel methods / machine learning
Date
Fri, 8 Apr 2011 15:58:31 +0100
I don't think there's much that's visible. I'll guess wildly that
there may be Stata stuff that is invisible to us, in the backrooms of
some companies or agencies. But that's an known unknown to me at most.
I think there are lots of people who would be interested in playing
with this in an environment in which there was also serious data
management, e.g. Stata. Most crime-solving TV stories would seem to
depend on something similar!
Nothing stops people writing code for this in Stata, except that to
write a good set of supporting packages would seem to be a substantial
undertaking. My main guess is that in practice almost everybody doing
this -- and doing it visibly -- is probably doing it using some other
language. I'd guess there is a whole heap more in MATLAB, for example.
I think this stuff divides the world in various senses. One is on a
small level. "ML" I take to mean "machine learning" in your posting
but a lot of people on this list are likely to think "maximum
likelihood".
Another is on a larger level. Although lots of very smart people are
involved, I get a sense of a search for a Holy Grail and a rather
fickle field in which what is top and what is not varies very rapidly.
I remember a time when people seemed to going round saying "neural
nets" all the time, and so forth, but now we have different buzzwords.
Reminds me of Bayesian stuff...
Nick
On Fri, Apr 8, 2011 at 3:37 PM, Diego Navarro <[email protected]> wrote:
> Are there any Stata packages in the works for kernel trick methods
> such as kernel PCA/kernel SVM?
>
> I know SVM isn't Stata's cup of tea, but I really need kernel PCA
> these days, and the jerry-rigged code I hacked together for one
> project is pretty brittle, doesn't interact well with other Stata
> commands, will break at missing data and ill-behaved formulas, etc.
> etc. (Besides, it's so messy that I don't know for a fact that it
> conforms to the idea of kernel PCA everyone else is expecting).
>
> I did try "search kernel, net" and such, but there are just too many
> results. To ask too wide a question, do you think there's a future in
> Stata for machine learning methods? I came to Stata from econometrics,
> but my work is increasingly "ML-enhanced econometrics", and while I
> learned R (which has the basic kernel trick packages, for example), I
> don't care much for its ad-hockish palette of CLOS-like object
> structures.
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/