Transformers cheat sheets

dunders glossary

Type	Dunder	Meaning	`sktime` class
compose	`*`	chaining/pipeline - also works with other estimator types	type dependent
compose	`**`	chaining to secondary input of another estimator	type dependent
compose	`+`	feature union	`FeatureUnion`
interface	`~`	invert	`InvertTransform`
structural	`¦`	multiplexing (“switch”)	type dependent
structural	`-`	optional passthrough (“on/off”)	`OptionalPassthrough`

delay fitting to transform via sktime.transformations.compose.FitInTransform
any pandas method via sktime.transformations.compose.adapt.PandasTransformAdaptor
date/time features via sktime.transformations.date.DateTimeFeatures
lags via transformations.lag.Lag
differences, first and n-th, via transformations.difference.Differencer
scaled logit via transformations.scaledlogit.ScaledLogitTransform

Common types of transformation in sktime:

from	to	base class	examples (sci)	examples (`sktime`)
time series	scalar features	`BaseTransformer` (`Primitives` output)	`tsfresh`, or 7-number-summary	`Catch22`, `SummaryTransformer`
time series	time series	`BaseTransformer` (`Series`, `instancewise`)	detrending, smoothing, filtering, lagging	`Detrender`, `Differencer`, `Lag`, `Filter`
time series panel	also a panel	`BaseTransformer` (`Series` output)	principal component projection	`PCATransformer`, `PaddingTransformer`
two feature vectors	a scalar	`BasePairwiseTransformer`	Euclidean distance, L1 distance	`ScipyDist`, `AggrDist`, `FlatDist`
two time series	a scalar	`BasePairwiseTransformerPanel`	DTW distance, alignment kernel	`DtwDist`, `EditDist`

first three = “time series transformers”, or, simply, “transformers”

all “transformers” follow the same base interface.

“pairwise transformers” have separate base interface (due to two inputs)

include distances and kernels between time series or feature vectors

all inherit BaseObject and follow unified skbase interface with get_params, get_fitted_params, etc

Generated using nbsphinx. The Jupyter notebook can be found here.