Quantitative and Qualitative Comparison of Decision Map Techniques for Explaining Classification Models

Abstract

Visualization techniques for understanding and explaining machine learning models have gained significant attention. One such technique is the decision map, which creates a 2D depiction of the decision behavior of classifiers trained on high-dimensional data. While several decision map techniques have been proposed recently, such as Decision Boundary Maps (DBMs), Supervised Decision Boundary Maps (SDBMs), and DeepView (DV), there is no framework for comprehensively evaluating and comparing these techniques. In this paper, we propose such a framework by combining quantitative metrics and qualitative assessment. We apply our framework to DBM, SDBM, and DV using a range of both synthetic and real-world classification techniques and datasets. Our results show that none of the evaluated decision-map techniques consistently outperforms the others in all measured aspects. Separately, our analysis exposes several previously unknown properties and limitations of decision-map techniques. To support practitioners, we also propose a workflow for selecting the most appropriate decision-map technique for given datasets, classifiers, and requirements of the application at hand.