CARE: coherent actionable recourse based on sound counterfactual explanations

Abstract

AbstractCounterfactual explanation (CE) is a popular post hoc interpretability approach that explains how to obtain an alternative outcome from a machine learning model by specifying minimum changes in the input. In line with this context, when the model’s inputs represent actual individuals, actionable recourse (AR) refers to a personalized CE that prescribes feasible changes according to an individual’s preferences. Hence, the quality of ARs highly depends on the soundness of underlying CEs and the proper incorporation of user preferences. To generate sound CEs, several data-level properties, such as proximity and connectedness, should be taken into account. Meanwhile, personalizing explanations demands fulfilling important user-level requirements, like coherency and actionability. The main obstacles to inclusive consideration of the stated properties are their associated modeling and computational complexity as well as the lack of a systematic approach for making a rigorous trade-off between them based on their importance. This paper introduces CARE, an explanation framework that addresses these challenges by formulating the properties as intuitive and computationally efficient objective functions, organized in a modular hierarchy and optimized using a multi-objective optimization algorithm. The devised modular hierarchy enables the arbitration and aggregation of various properties as well as the generation of CEs and AR by choice. CARE involves individuals through a flexible language for defining preferences, facilitates their choice by recommending multiple ARs, and guides their action steps toward their desired outcome. CARE is a model-agnostic approach for explaining any multi-class classification and regression model in mixed-feature tabular settings. We demonstrate the efficacy of our framework through several validation and benchmark experiments on standard data sets and black box models.