[论文解读] Customer Churn Prediction Model using Explainable Machine Learning

Jitendra Maan, Harsh Maan|arXiv (Cornell University)|Mar 2, 2023

Customer churn and segmentation被引用 8

一句话总结

论文使用可解释的 ML 构建客户流失预测模型，确定 XGBoost 为最有效的分类器，并提出基于 Shapley 值的特征重要性解释方法。

ABSTRACT

It becomes a significant challenge to predict customer behavior and retain an existing customer with the rapid growth of digitization which opens up more opportunities for customers to choose from subscription-based products and services model. Since the cost of acquiring a new customer is five-times higher than retaining an existing customer, henceforth, there is a need to address the customer churn problem which is a major threat across the Industries. Considering direct impact on revenues, companies identify the factors that increases the customer churn rate. Here, key objective of the paper is to develop a unique Customer churn prediction model which can help to predict potential customers who are most likely to churn and such early warnings can help to take corrective measures to retain them. Here, we evaluated and analyzed the performance of various tree-based machine learning approaches and algorithms and identified the Extreme Gradient Boosting XGBOOST Classifier as the most optimal solution to Customer churn problem. To deal with such real-world problems, Paper emphasize the Model interpretability which is an important metric to help customers to understand how Churn Prediction Model is making predictions. In order to improve Model explainability and transparency, paper proposed a novel approach to calculate Shapley values for possible combination of features to explain which features are the most important/relevant features for a model to become highly interpretable, transparent and explainable to potential customers.

研究动机与目标

• 应对数字化订阅制市场中预测与保留客户的挑战。
• 评估基于树的机器学习模型在流失预测中的表现，以识别最有效的方法。
• 强调模型可解释性，以向相关方解释预测。
• 提出一种基于 Shapley 值 的新方法来解释特征贡献并提升透明度。

提出的方法

• 比较用于流失预测的树基机器学习方法。
• 在评估的模型中识别 Extreme Gradient Boosting (XGBoost) 作为最优分类器。
• 开发一种新颖的方法来计算特征组合的 Shapley 值，以解释模型预测。
• 聚焦可解释性与透明度，帮助相关方理解流失预测。

实验结果

研究问题

• RQ1 在树基方法中，哪种机器学习模型在流失预测方面表现最佳？
• RQ2 如何提高对潜在客户的流失预测的可解释性与透明度？
• RQ3 根据基于 Shapley 值 的解释，哪些特征在预测流失方面最具影响力？
• RQ4 基于 Shapley 值 的方法能否有效解释流失模型中的特征组合？

主要发现

• XGBoost 分类器被确定为流失预测的最优解决方案。
• 提出一种新的基于 Shapley 值 的方法来解释可能的特征组合对预测的影响。
• 研究强调模型可解释性，使流失预测对客户和相关方更易理解。