Hi Nelle,
Thanks for your answer! In the MDS code, the precision is hard-wired to:
res = 100 * np.finfo(np.float).resolution
So I cannot change the precision without modifying the source. If I do
modify the source code, and lower the precision to 1e-8, I'm able to run
MDS successfully (yay!).
As I mentioned in my email, my distance matrix was computed with sklearn,
which is precise up to 1e-9. I think the problem lies on MDS expecting
unrealistic high precisions in the distance matrix.
Best,
Jose
--
José P. González-Brenes, Ph.D.
www.josepablogonzalez.com
On Wed, Dec 11, 2013 at 6:13 PM, Nelle Varoquaux
<[email protected]>wrote:
> On 11 December 2013 23:30, José Pablo González <[email protected]>
> wrote:
> > Hello,
> >
> > I am new in sklearn, but so far I'm impressed. I think it is a great
> > library!
> >
> > I think I found a little bug in the Multidimensional Scaling Function. It
> > expects "perfect" similarity matrices, which it's often not possible with
> > real data. Moreover, sklearn's functions are unable to get the expected
> > precision.
> >
> > For example if I use sklearn to computea similarity matrix of euclidean
> > distances:
> >
> > similarities = euclidean_distances(data)
> > print np.abs(similarities - similarities.T).max()
> > # this outputs 5.26835606386e-09, which I'd it's precise enough
> >
> >
> > Unfortunately, when I try to calculate MDS:
> > mds = manifold.MDS(n_components=2, dissimilarity="precomputed", n_jobs=1)
> > pos = mds.fit(similarities)
> >
> > I get a precision error. I have already tried casting to float64 as
> > suggested here, but it doesn't work either. There's no way to change the
> > precision of this, only the convergence precision.
> >
> > Am I overlooking at something?
>
> Nope, this is indeed a problem from the MDS code. You can provide a
> fix for this by raising the precision required.
> Also, depending on the algorithm used to solve the MDS, a symmetric
> matrix should or shouldn't be required. I think in the case of the
> SMACOF algorithm (currently used in sklearn), it isn't even required.
> Someone should have a look at the litterature to check this (I should
> be able to in a couple of weeks).
>
> Cheers,
> N
>
> >
> > Thanks,
> > Jose
> >
> >
> > --
> > José P. González-Brenes, Ph.D.
> > www.josepablogonzalez.com
> >
> >
> >
> >
> ------------------------------------------------------------------------------
> > Rapidly troubleshoot problems before they affect your business. Most IT
> > organizations don't have a clear picture of how application performance
> > affects their revenue. With AppDynamics, you get 100% visibility into
> your
> > Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics
> > Pro!
> >
> http://pubads.g.doubleclick.net/gampad/clk?id=84349831&iu=/4140/ostg.clktrk
> > _______________________________________________
> > Scikit-learn-general mailing list
> > [email protected]
> > https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
> >
>
>
> ------------------------------------------------------------------------------
> Rapidly troubleshoot problems before they affect your business. Most IT
> organizations don't have a clear picture of how application performance
> affects their revenue. With AppDynamics, you get 100% visibility into your
> Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics
> Pro!
> http://pubads.g.doubleclick.net/gampad/clk?id=84349831&iu=/4140/ostg.clktrk
> _______________________________________________
> Scikit-learn-general mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>
------------------------------------------------------------------------------
Rapidly troubleshoot problems before they affect your business. Most IT
organizations don't have a clear picture of how application performance
affects their revenue. With AppDynamics, you get 100% visibility into your
Java,.NET, & PHP application. Start your 15-day FREE TRIAL of AppDynamics Pro!
http://pubads.g.doubleclick.net/gampad/clk?id=84349831&iu=/4140/ostg.clktrk
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
