New Featurizers (removing old, ineffective ones)

[ardunn]

Our results depend entirely on the quality of our descriptors. We need to add better ones, if possible, and prune the "best" sets to keep only the best ones.

A few ideas for featurizers to add to matminer (some of which are already issues on the matminer repo):

• SOAP thru dscribe or native implementation
• MBTR thru dscribe or native
• U-apI thru ABCNN or native
• Matscholar word embeddings (for definite compositions, structure prototypes, and combinations of these via dot product)
• A structure featurizer based on: A site featurizer that averages/combines composition properties, where the compositions are the bonds [determined via crystalnn], and the composition properties can be from any combination of composition featurizers (this is already done somewhat through Ward et al. LocalPropertyDifference)
• Flat version of SineCoulombMatrix or CoulombMatrix

Adding these and then evaluating whether we should add them to automatminer (i.e., by evaluating them by themselves on some easy test problems using the same model) should be a priority.

[ardunn]

update to this: tried raw word embeddings for various compositions, structure types, and combinations of them (via same method as the paper): doesn't work so well

[ardunn]

update: flat cm updated thru matminer #367
performance on elastic tensor is not great, but fast and few/dense features returned, so that is good

[ardunn]

update: added soap thru matminer #370

[ardunn]

Consider adding a featurizer which searches MP for close stoichiometries and determines the energy with respect to the hull by just drawing a line between its neighbors etc. (some problems such as never predicting an e being on a hull but still might be useful)