Functional relevance based on the continuous Shapley value

  • 2024-11-27 18:20:00
  • Pedro Delicado, Cristian Pachón-García
  • 0

Abstract

The presence of Artificial Intelligence (AI) in our society is increasing,which brings with it the need to understand the behaviour of AI mechanisms,including machine learning predictive algorithms fed with tabular data, text,or images, among other types of data. This work focuses on interpretability ofpredictive models based on functional data. Designing interpretability methodsfor functional data models implies working with a set of features whose size isinfinite. In the context of scalar on function regression, we propose aninterpretability method based on the Shapley value for continuous games, amathematical formulation that allows to fairly distribute a global payoff amonga continuous set players. The method is illustrated through a set ofexperiments with simulated and real data sets. The open source Python packageShapleyFDA is also presented.

 

Quick Read (beta)

loading the full paper ...