基于Stacking方法的银行客户产品认购预测

Bank Customer Product Subscription Prediction Based on the Stacking Method

摘要: 本文通过对银行客户数据的挖掘与建模，旨在预测客户是否会购买银行产品。采用融合随机森林、LightGBM、XGBoost及多层感知机的Stacking集成学习方法，先以四种模型作为基学习器挖掘数据中线性、非线性及复杂特征模式，再通过逻辑回归元学习器整合优化预测结果。实验结果显示，该集成模型预测准确性显著优于单一模型，在客户产品认购行为预测任务中表现出色。在应用中，基于高认购概率模型输出的重要特征与客户行为标签完成6个客群的划分，形成多维度用户画像体系，为精准营销与客户关系管理提供支持。

Abstract: This study aims to predict whether customers will purchase bank products through data mining and modeling of bank customer data. A Stacking ensemble learning method integrating Random Forest, LightGBM, XGBoost, and Multilayer Perceptron (MLP) is adopted: firstly, these four algorithms serve as base learners to explore linear, non-linear, and complex feature patterns in the data, and then a Logistic Regression meta-learner is employed to integrate and optimize the prediction results. Experimental results demonstrate that the proposed ensemble model significantly outperforms single models in prediction accuracy and performs excellently in the task of forecasting customers’ product subscription behaviors. In the application phase, six customer groups are divided based on the key features output by the high subscription probability model and customer behavior labels, and a multi-dimensional user portrait system is established, thereby providing robust support for precision marketing and customer relationship management.