Methods for evaluating attribution methods for generating relevance maps for a machine learning model

Abstract

The invention relates to a method (100) for evaluating attribution methods for generating relevance maps for a machine learning model, comprising: - generating (101) a first relevance map for at least one image using the machine learning model based on a first attribution method, - generating (102) a second relevance map for the at least one image using the machine learning model based on a second attribution method, - analyzing (103) the generated first relevance map and the generated second relevance map, - evaluating (104) the first and the second attribution methods based on a result of the analysis (103), wherein the first and the second attribution methods each determine in a different way an influence of individual features of the at least one image on a model prediction of the machine learning model based on the at least one image and represent this by the respective generated relevance map.

Images