On Wed, Jan 22, 2014 at 9:48 AM, Mathieu Blondel <[email protected]>wrote:
>
> Something I was wondering is whether sparse support in decision trees
> would actually be useful. Do decision trees (or ensembles of them like
> random forests) work better than linear models for high-dimensional data?
>
I share your point of view.
>
> It would be nice to take the News20 dataset, pre-select the top 10k
> features (or more if possible) then measure test accuracy on the densified
> dataset. I would be very interested in hearing the results.
>
I would also be interested in it. Applying random projection down to 100
<-> 1000 features would be also very interesting.
I'm thinking also to features hashing with multiple probes that I think
it's not currently implemented in sklearn.
The idea could be that of mapping via murmurhash "literal" features or
features id to 128 bit, group the hash in 16 group of 8 bits.
each group is a probe on the projected 256 features (8 bits). the pattern
of the probe would be -1, 1, -1, 1, ....
As anyone ever performed such experiments?
Paolo
------------------------------------------------------------------------------
WatchGuard Dimension instantly turns raw network data into actionable
security intelligence. It gives you real-time visual feedback on key
security issues and trends. Skip the complicated setup - simply import
a virtual appliance and go from zero to informed in seconds.
http://pubads.g.doubleclick.net/gampad/clk?id=123612991&iu=/4140/ostg.clktrk
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
