# Counterfactual Shapley Additive Explanations

**Authors:** Emanuele Albini, Jason Long, Danial Dervovic, Daniele Magazzeni  
**arXiv:** [2110.14270](https://arxiv.org/abs/2110.14270)  
**Date:** 2021-10-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2110.14270v4/

## Key points

- The paper discusses the use of feature attributions in model explanations and their application in the actionable recourse setting
- The authors propose a variant of SHAP called Counterfactual SHAP (CF-SHAP) that incorporates counterfactual information
- Careful consideration of background datasets is highlighted when using Shapley values for feature attributions in the actionable recourse setting
- CF-SHAP outperforms existing methods when evaluated on public datasets using tree ensembles, as demonstrated by the "counterfactual-ability" score
- Future research directions include exploring alternative notions of action and cost functions based on psychology findings, testing the approach on different models like neural networks, and incorporating potentially model-agnostic counterfactual explanations
- Overall, CF-SHAP aims to strengthen and clarify the link between actionable recourse and feature attributions while improving outcomes for model consumers.

## Summary

The paper discusses the use of feature attributions in model explanations and their application in the actionable recourse setting, where the goal is to improve outcomes for model consumers. The authors propose a variant of SHAP (SHapley Additive exPlanations), called Counterfactual SHAP (CF-SHAP), which incorporates counterfactual information to produce a background dataset for use within the marginal Shapley value framework. The authors highlight the need for careful consideration of background datasets when using Shapley values for feature attributions in the actionable recourse setting. They demonstrate the efficacy of CF-SHAP by introducing a quantitative score called "counterfactual-ability" and show that CF-SHAP outperforms existing methods when evaluated on public datasets using tree ensembles. In addition, they suggest several directions for future research. They propose exploring alternative notions of action and cost functions based on findings in psychology regarding how users interpret feature attributions and how they change their behavior accordingly. They also suggest testing their approach on different models, such as neural networks, and incorporating potentially model-agnostic counterfactual explanations. Overall, this work aims to strengthen and clarify the link between actionable recourse and feature attributions by proposing CF-SHAP as an improved method for generating feature attributions with counterfactual information. The proposed method shows promise in improving outcomes for model consumers and provides a quantitative metric for evaluating feature attributions.