Explainable AI Metrics and Properties for Evaluation and Analysis of Counterfactual Explanations : Explainable AI Metrics and Properties for Evaluation and Analysis of Counterfactual Explanations

• Main Author: Singh, Vandita
• Format: Others
• Language: English
• Published: Uppsala universitet, Institutionen för informationsteknologi 2021
• Subjects: Engineering and Technology; Teknik och teknologier
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-462552

With the recent surge of state-of-the-art AI systems, there has been a growing need to provide explanations for these systems to ensure transparency and understandability by the stakeholders. This has directed the development of increasingly popular Explainable AI systems applicable to a wide range of applications. This trend designates the need for the ability to quantitatively assess and analyze the theoretical and behavioral characteristics of these Explainable AI systems. This study aimed at the following (i) to identify metrics and properties applicable to post-hoc counterfactual explanation methods (a mechanism for generating explanations), (ii) assess the applicability of the identified metrics and properties to compare counterfactual examples, and (iii) analyze the functional and operational characteristics of the selected counterfactual explanation methods. A proof-of-concept tool was designed to perform the given task. A pipeline was implemented, which comprised of selecting a data set, training some suitable classifier, deploying counterfactual generation method, and performing analysis using the metrics selected for implementation. Four counterfactual explanation methods were selected and analyzed against the metrics - distance measures, loss measures, change score values, recourse values in terms of the value of features, and neighborhood. It was found that the metrics provided a means to analyze the characteristics of the selected counterfactual explanation methods. It could also be reasoned which explainers generated better counterfactual explanations for a particular instance by comparing the metrics across different explainers for the same instance. Also, insight could be gained to understand whether a particular counterfactual instance was useful enough or not by comparing the implemented metrics concerning an instance of the desired class. A fact sheet enlisting the functional and operational characteristics of the selected counterfactual methods has been presented. The scope of the work was limited to data sets being comprised of tabular data and images, simpler classification methods were chosen to be incorporated into the pipeline, and the type of explanations was confined to counterfactual explanations.