基于Firth逻辑回归与随机森林的在线购买意向预测

Predicting Online Purchase Intention via Firth Logistic Regression and Random Forest

摘要: 在数字经济由规模扩张向质量驱动转型的背景下，预测用户购物意图对提升电商转化率至关重要。本研究利用UCI在线会话数据集作为基准测试样本，旨在验证Firth逻辑回归与随机森林模型在处理类别不平衡及数据稀疏性问题上的预测效能，为实时意图预测提供算法框架参考。研究发现：第一，实时行为是电商转化的核心指标，其中页面价值对购买概率有极强的正向驱动作用，尽管其存在统计内生性，但在动态识别逻辑下是捕捉访客从“浏览”转向“决策”阶段的关键先导指标，而退出率则产生显著的负向抑制效应；第二，外部情境具有显著调节作用，大促月份通过激活消费者的价格预期与限时心理，明显提升购买概率；第三，新老访客存在决策异质性，回访者对页面流畅度、周末窗口更为敏感，而新访客则更依赖高价值页面的即时引导。本文为电商企业动态识别高价值会话、进行智能化干预及实施差异化运营提供了指导性的算法支撑。

Abstract: As the digital economy transforms from scale expansion to quality-driven growth, it is significant to predict users’ shopping intention to enhance the conversion rates of e-commerce. This study uses the UCI online session dataset as a benchmark sample to verify the prediction efficiency of Firth Logistic Regression and Random Forest models in dealing with the problems of category imbalance and data sparsity, and provides an algorithm framework for real-time intention prediction. The findings reveal that: (1) Real-time behavior is the primary indicator of e-commerce transformation, in which page values have a strong positive driving effect on purchase probability. Despite its inherent statistical endogeneity, it serves as a critical precursor signal for capturing the transition from the “browsing” to the “decision-making” stage within dynamic recognition logic. Conversely, exit rates have a significant negative inhibitory effect; (2) External contexts play a significant moderating role. In big promotion months, the purchase probability is significantly improved by activating consumers’ price expectation and limited-time psychological triggers; (3) Significant decision heterogeneity exists between new and returning visitors. Old visitors are more sensitive to page fluency and weekend windows, while new visitors rely more on the real-time guidance of high-value pages. This study provides guiding algorithm decision support for e-commerce enterprises to dynamically identify high-value sessions, carry out intelligent intervention, and implement differentiated operational strategies.