Contextual Shift Method (CSM)

TL;DR: This study addresses the limitations of Explainable AI methods by introducing the Contextual Shift Method (CSM), which generates meaningful, realistic artificial data points, reducing the creation of points in low data density areas and improving model interpretability.

Abstract: Explainable AI approaches often create artificial data points to test a given model. Sometimes the created data points are located in areas with low data density, and they are unlikely or even impossible combinations of values. Hence, interpreting the model at those artificial points does not give trustworthy information. This becomes even more relevant the higher the dimensionality of the data. We examine the challenges of creating meaningful, realistic data points, which are essential for many Explainable AI methods. Based on this knowledge, we define a contextual shift as a meaningful artificial data point. The problem of not generating contextual shifts is true for the quantile shift method. We propose the Contextual Shift Method (CSM), which improves the quantile shift method by generating contextual shifts. We show that the CSM reduces the amount of data points created in low data density areas.