电商运输准时性的可解释预测：基于BO与集成学习

Explainable On-Time Delivery Prediction for E-Commerce: A BO and Ensemble Learning Approach

摘要: 随着电子商务的快速发展，运输准时性已成为影响客户体验与企业竞争力的核心因素。然而，传统预测方法难以应对物流环节中的复杂不确定性。为此，本文构建了一个融合贝叶斯优化、Stacking集成学习与SHAP可解释性分析的运输准时性预测框架。首先，基于Kaggle公开电商数据集进行特征工程；随后，应用贝叶斯优化对LightGBM、RF等多种机器学习模型进行超参数调优，并以Stacking方法集成，以提升对延迟订单的识别能力。实验结果表明，经贝叶斯优化后，模型平均召回率提升约7%，而集成模型取得了最优的综合性能，召回率达到73.64%。SHAP可解释性分析进一步揭示，折扣率和商品重量是影响模型输出的两个最强有力的驱动因素。本研究不仅提供了一套现实可行的预测框架，更通过可解释分析为电商物流的主动时效管理与精准干预提供了决策依据。

Abstract: With the rapid development of e-commerce, on-time delivery has become a critical factor influencing customer experience and corporate competitiveness. However, traditional prediction methods struggle to address the complex uncertainties inherent in logistics processes. To tackle this, this paper constructs a transportation timeliness prediction framework that integrates Bayesian optimization, Stacking ensemble learning, and SHAP interpretability analysis. First, feature engineering is performed based on a publicly available e-commerce dataset from Kaggle. Subsequently, Bayesian optimization is applied to fine-tune the hyperparameters of multiple machine learning models, including LightGBM and Random Forest, and these models are integrated using the Stacking method to enhance the identification capability for delayed orders. Experimental results show that after Bayesian optimization, the average recall rate of the models improved by approximately 7%, and the ensemble model achieved the best overall performance, with a recall rate of 73.64%. SHAP interpretability analysis further reveals that the discount rate and product weight are the two most powerful driving factors influencing model output. This study not only provides a practical and feasible prediction framework but also offers decision-making support for proactive timeliness management and precise intervention in e-commerce logistics through interpretable analysis.