# "How to make them stay?" -- Diverse Counterfactual Explanations of Employee Attrition

**Authors:** André Artelt, Andreas Gregoriades  
**arXiv:** [2303.04579](https://arxiv.org/abs/2303.04579)  
**Date:** 2023-03-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.04579v1/

## Key points

- Employee attrition can negatively impact an organization's competitiveness and performance
- Machine learning has been used to predict employee attrition and suggest proactive measures
- However, the best-performing ML methods are black box techniques that cannot be easily interpreted
- Counterfactual explanation methods have been proposed to address this issue, but current methods focus on individual cases rather than groups of employees
- The authors propose a new method utilizing counterfactual explanations focusing on multiple attrition cases from historical data to identify specific business policy changes that can revert attrition of specific employees by providing actionable recommendations based on multiple instances rather than individual employees who left the company.
- This approach provides comprehensive HR policies that predict turnover intention and reduces negative effects on business performance and competitiveness.
- The proposed method goes beyond other interpretable ML methods such as SHAP by providing specific policy changes optimized on multiple instances.
- Future work will focus on providing weights to features manipulated so that the method can provide ranked recommendations based on how easy they can be implemented.

## Summary

Employee attrition is a major challenge for organizations as it can directly impact their competitiveness and performance. The cost and time required to attract and retain talented employees make it crucial for businesses to increase employee retention rates. Machine learning (ML) has been applied in various aspects of human resource management, including attrition prediction, to provide insights on proactive measures that can be taken to prevent talented employees from quitting. However, the best-performing ML methods, such as ensemble or deep neural networks, are black box techniques that cannot be easily interpreted. To address this issue, several explainability frameworks have been proposed, including counterfactual explanation methods that can explain and recommend actions to obtain the desired outcome. However, current counterfactual explanations methods focus on optimizing changes made on individual cases to achieve the desired outcome. In contrast, in the case of employee attrition prevention, it is important to identify the optimum interventions that an organization needs to make to its practices/policies to prevent or minimize attrition probability for a group of employees. In this paper, the authors propose a new method utilizing counterfactual explanations focusing on multiple attrition cases from historical data. This method identifies specific business policy changes that can revert attrition of specific employees by providing actionable recommendations based on multiple instances rather than individual employees who left the company. This approach provides comprehensive HR policies that predict turnover intention and reduces negative effects on business performance and competitiveness. The proposed method goes beyond other interpretable ML methods such as SHAP by providing specific policy changes optimized on multiple instances (e.g., employees in a particular department whose skills cannot easily be replenished). It explains what management needs to do to prevent attrition of these employees by providing actionable business recommendations. Overall, this work highlights the importance of using multiple instances when generating recommendations for preventing employee attrition rather than relying solely on individual cases. Future work will focus on providing weights to features manipulated so that the method can provide ranked recommendations based on how easy they can be implemented. The method will also be extended to eliminate automatically infeasible solutions by enabling the user to provide a range of feasible feature values and comparing results with similar work under different scenarios.