Block model

[emorice]

• [x] Actual non-linear test case
• [x] GMM labeling
• ~[ ] Block RWSS~
• ~[ ] Pre-scaled version~
• ~[ ] Post-scaling ?~

[emorice]

Test case:

• [x] generate from spectrum
• [x] bend it

[emorice]

Status: the dataset written for that (case nonlinear) is very much in use.

Sklearn GMM has been integrated as model.gmm. Early testing showed that it's not great at finding useful groups, so we branched out to develop model.igmm.

Predictions have been done too, but we've departed from the idea of a block RWSS. Indeed, IGMM has been implemented with soft weights (both out of principle, but also because it allows to plug in continuous optimization and bypass the convergence problems seen in early CMK prototypes that used hard weights). From there, it was very natural to write the prediction in the same framework, and hard to fit back soft weights into the original RWSS design.

Therefore, the idea of a native large-dim group model has been essentially cancelled as the native low-dim can do for now.