Encoders

[krystophny]

• encoders are currently a bit messy: input/output encoders and special cases to handle decoding the prediction variance

Split Encoders ?

• raw encoder objects which just know encode and decode and act on arrays
• higher level objects which are aware of the underlying structure
  • labelled columns / variables - e.g. xarray
  • encode_input, decode_variance etc
• alternatively let this raw transformation be class methods of the encoder objects → can be used without instantiating objects

[Rykath]

Discussion on 2022-03-18

• decoding the variance only works well with linear transformations (Normalization, PCA)

  • Could add method to these transformations and use it in surrogates. If it doesn't exist, predict should yield exception.

• for non linear transformations it makes more sense to decode the confidence bounds

  • For plotting routines, surrogate should provide confidence bands directly as output instead of variance
  • Optional: Add term based on derivative of transformation

[Rykath]

For dimensionality reduction (PCA, KL) there should be an option to explicitly set the number of output dimensions / principal values as an alternative to the tolerance.