MDS for Configurator footprint only with randomly sampled configurations

[mlindauer]

Please train the MDS mapping only on randomly sampled configurations (not on the ones optimized wrt EI). Of course, please still plot all configurations.

(As a reminder of what we discussed last week).

[shukon]

sklearns MDS does not provide a transform-method, though there is an open PR: https://github.com/scikit-learn/scikit-learn/pull/9834 I could try to work on the open PR in sklearn to fix this, but I will first look for another implementation/ an alternative algorithm, since I never commited to sklearn before.

[shukon]

So looking further into this yielded a more fundamental problem. The MDS-algorithm reduces dimensionality by shifting points around iteratively until the distances show more or less the same distance-matrix as in the higher-dimensional space. Inserting points afterwards is not easily feasible since we have to shift points again. We could, maybe, use a Kernel-function to approximate the mapping (following this idea).

[mlindauer]

We could, maybe, use a Kernel-function to approximate the mapping (following this idea).

Sounds interesting. Is an implementation available? If not, how much time do you expect you will need to implement it?

[jsoutherland]

@shukon that paper mentioned was the basis for https://github.com/scikit-learn/scikit-learn/pull/9834 by @webdrone which switches the method of computing MDS (no longer using smacof). You may find the code in that PR to be useful.