Involved tasks

FEDOT is currently capable of solving:

• Classification

• Regression

• Time-series forecasting ([uni/multi]variate)

Pipeline building

FEDOT uses open-source library named GOLEM for optimization and learning of graph-based pipelines with meta-heuristic methods.

The library is potentially applicable to any graph-based optimization problem with clearly defined fitness function on it.

Sure enough, you may use your own custom optimization algorithms, see Automated Design of Pipelines.

Data preprocessing

FEDOT uses two types of preprocessing: obligatory and optional.

Note

Preprocessing is optional (see use_input_preprocessing main API parameter), so you can save some time if your dataset (input data) is already preprocessed.

Obligatory preprocessing, as you might guess, solves major problems that can disrupt or complicate data processing, such as:

Obligatory preprocessing

Problem	Solution
‘inf’ values in features	replace
huge amount of nans in features or targets	drop
binary categorical form of the features or targets	binarize
extra spaces in categorical features	trim

Optional preprocessing depends on a composed pipeline structure, and is applied only if it is necessary for a next model from a processing queue to work.

Optional preprocessing

Problem	Solution
nans in features	impute
non-binary categorical features	LabelEncode or OneHotEncode

But depending on the pipeline structure, it might be ommited:

See also

Extended explanation

Note

Both obligatory and optional preprocessing are applied only once.