Investigating Desirable Properties of Inverse Projections and Decision Maps

Abstract

Inverse projection techniques enable the creation of decision maps which help the visual exploration of trained classification models. However, different inverse projections lead to significantly different decision maps for the same model, leading to uncertainty in their interpretation. Recent work compared three inverse projection techniques from the perspective of their intrinsic dimensionality and showed that all three techniques visualize only two-dimensional substructures in the data space. We extend this evaluation in several directions. First, we consider three additional inverse projections thereby covering, to our knowledge, all such techniques in existence. Secondly, we correlate the quality of the inverse projections with their ability to depict certain types of data structures. Finally, we study the smoothness of the structures created by inverse projections. Our results show that all inverse projection techniques essentially cover only two-dimensional structures in the data space and that the smoothness of such structures is inversely correlated with their ability to approximate data points. Based on our findings, we also propose ways to select inverse projections which lead to interpretable decision maps.