Regional explanation according to us

Tools already exist to explain a model in a « global » way. PDP, or Partial Dependence Plots, is perhaps the most widespread.

The SHAP library (which implements Shapley additive values from game theory) introduced local explanation.

SHAP values

Global or local, these approaches are not « at the right granularity » in our opinion. Only an intermediate approach, or regional one, allows for an understanding by the business of the correlations revealed by the model.

Validated during the ACPR tech sprint on explainability (the 3 laureates, of which we were part, proposed such a regional approach), this conviction was the starting point for the development of our product AntakIA: substituting the original model with simpler, explainable, and frugal models within regions:

Therefore, we replace the black box with a mosaic of simpler models. This image of the mosaic gave its name to our product: the ancient city of Antioch was famous for its mosaics – the modern name of this now Turkish city is AntakIA 🙂

Et c’est aussi l’acronyme en anglais de « A New Tool To Acquire Knowledge from AI« .

A noter que notre approche s’inscrit parfaitement dans la méthodologie que Descartes formalisait dès 1637 dans son « Discours de la méthode » !

René Descartes