XPC : eXplainability through Positive Contributions

Note: code is currently undergoing rework.

The package enables to decompose the forecasts of a model according to specified parts, e.g a climate and non-climate part.

The method is based on Shapley values in addition to a novel pipeline to aggregate, heighten and normalize. We obtain an interpretable decomposition of the time series.

More details will be provided in an upcoming paper. Example notebooks are available to experiment on the package.

Author : Gaspard Berthelier.

Developed during my end-of-studies intership in 2024, supervised by Margaux Brégère, within the OSIRIS Departement (R39) at EDF.

Logo by Anna Bondarenko.