eduzhai > Applied Sciences > Computer Science >

Explaining black box decisions by Shapley cohort refinement

  • KanKan
  • (0) Download
  • 20210425
  • Save

... pages left unread,continue reading

Document pages: 26 pages

Abstract: We introduce a variable importance measure to quantify the impact ofindividual input variables to a black box function. Our measure is based on theShapley value from cooperative game theory. Many measures of variableimportance operate by changing some predictor values with others held fixed,potentially creating unlikely or even logically impossible combinations. Ourcohort Shapley measure uses only observed data points. Instead of changing thevalue of a predictor we include or exclude subjects similar to the targetsubject on that predictor to form a similarity cohort. Then we apply Shapleyvalue to the cohort averages. We connect variable importance measures fromexplainable AI to function decompositions from global sensitivity analysis. Weintroduce a squared cohort Shapley value that splits previously studied Shapleyeffects over subjects, consistent with a Shapley axiom.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...