Sanity Check for Training Instance Influences for Prediction

Abstract

Showing "training instances that contribute to the prediction" as the reason for the prediction of the machine learning model can improve users' satisfaction. As a measurement of "the contribution of the training instance on the prediction", various methods have been proposed: "similarity of input between test and training instances", "the change in prediction when the training instance is excluded from training data" or "the cosine similarity of a gradient vector". In this study, we considered some requirements that contribution measurements should meet, and examined whether each measurement satisfies the requirements. We show that some measurements do not meet the requirements.