基于HGBT的电商客户流失预测与精细化营销策略

E-Commerce Customer Churn Prediction and Refined Marketing Strategies Using HGBT

摘要: 在电商市场竞争日益激烈的当下，提升用户留存对电商平台的稳健发展至关重要，预测用户流失并制定针对性营销策略也因此极具现实意义。本文以Retail Rocket零售电商平台数据为基础，构建了“模型预测–特征解释–策略落地”一体化融合框架。通过特征集重要性与边际增益分析解释流失驱动因素：新近度是核心特征，贡献占比达57.22%，频次次之，占比32.98%；不同风险层级下的主要特征存在明显的非线性阈值效应，例如高风险层“最小购买间隔”阈值为10.94天，同时表明“会话新近度 × 历史购买次数均值”等关键特征的交互关系。基于此，本文构建了“风险–特征–干预”分层策略，形成从流失预警到精准营销的完整闭环，论证了HGBT模型在处理电商行为序列数据时，应对非线性与交互效应相关挑战的独特优势，可为电商平台提供高效、可靠的决策支撑。

Abstract: In the current context of increasingly fierce competition in the e-commerce market, improving user retention is crucial for the stable development of e-commerce platforms, and thus predicting user churn and formulating targeted marketing strategies is of great practical significance. Based on the data from the Retail Rocket retail e-commerce platform, this paper constructs an integrated framework of “model prediction - feature interpretation - strategy implementation”. By analyzing feature set importance and marginal gain, the drivers of user churn are explained: recency is the core feature, accounting for 57.22% of the contribution, followed by frequency, accounting for 32.98%; the main features at different risk levels have obvious nonlinear threshold effects, for example, the threshold of “minimum purchase interval” in the high-risk layer is 10.94 days, and it also shows the interaction relationship of key features such as “session recency × average historical purchase times”. Based on this, this paper constructs a “risk-feature-intervention” hierarchical strategy, forming a complete closed loop from churn early warning to precise marketing, demonstrating the unique advantages of the HGBT model in dealing with challenges related to nonlinearity and interaction effects when processing e-commerce behavior sequence data, and can provide efficient and reliable decision support for e-commerce platforms.