missing features for prediction

ppdebreuck / modnet

MODNet: a framework for machine learning materials properties

MIT License

76 stars 32 forks source link

missing features for prediction #34

Open ppdebreuck opened 3 years ago

ppdebreuck commented 3 years ago

If some features are missing (from the optimal ones), when featurizing novel compounds (for predicting stage), predict() will not work. This happens for instance when some elements are not present in this prediction set, and corresponding features are thus not generated, but could have been selected during training.

ppdebreuck commented 1 year ago

Issue still present - can now be elegantly fixed by using (or adapting) our Imputer to remember and impute missing features needed for prediction. Sidenotes:

It would be better though that non-existing features (such as Miedema statistics) are fixed at the root (either update matminer or not use at all).
Imputation might be feature dependent... case (i) as above: unknown: replace by mean, -1,... Case (ii): missing but known: e.g. Element Fraction of missing element: obviously 0.

ppdebreuck commented 1 year ago

Tagging @gbrunin as partially involved in this.