Customer churn can be defined as the phenomenon of customers who discontinue their relationship with a company. This problem is transversal to many industries, including the software industry. This study uses Machine Learning to build a predictive model to identify potential churners in a Portuguese software house. Six popular Machine Learning models: Random Forest, AdaBoost, Gradient Boosting Machine, Multilayer Perceptron Classifier, XGBoost, and Logistic Regression, were developed to assess which one would have a better performance. The experimental results show that boosting techniques such as XGBoost present the best predictive performance. The XGBoost model presents a Recall of 0.85 and a ROC AUC of 0.86. Additionally to the model performance, the study of the model's feature importance revealed that some factors, such as the time to solve a support ticket, the type of application, the license age, and the number of incidents, significantly influence customer churn. These insights can help the software industry key drivers of churn and prioritize retention efforts accordingly.

客户流失可以定义为客户终止与公司的关系的现象。这个问题涉及许多行业，包括软件行业。本研究使用机器学习构建预测模型来识别葡萄牙软件公司中潜在的流失者。开发了六种流行的机器学习模型：随机森林、AdaBoost、梯度提升机、多层感知器分类器、XGBoost 和逻辑回归，以评估哪一种模型具有更好的性能。实验结果表明，XGBoost 等增强技术呈现出最佳的预测性能。XGBoost 模型的召回率为 0.85，ROC AUC 为 0.86。除了模型性能之外，对模型功能重要性的研究表明，一些因素（例如解决支持请求的时间、应用程序类型、许可证期限和事件数量）会显着影响客户流失。这些见解可以帮助软件行业确定流失的关键驱动因素，并相应地优先考虑保留工作。