电商用户流失预测训练模型

本模型基于SM4加密处理的用户脱敏数据（含交易/行为/投诉信息），提取近30天访问频次、消费间隔、消费金额波动率等核心特征。采用随机森林模型构建用户流失预测规则：通过多棵决策树投票机制（特征分裂阈值示例：消费金额＜100元且15天无访问行为）生成特征重要性排序，同步计算AUC值（ROC曲线下面积，评估整体排序能力）和召回率（召回率=被正确识别的流失用户/实际流失用户）。经交叉验证后，模型达到AUC≥0.85、召回率＞80%的标准，输出流失预警名单及核心规则（如：近30天未访问且消费额下降50%+无售后互动），支撑定向发放优惠券、调整商品推荐策略等运营动作。

This model is developed based on user anonymized data processed with SM4 encryption, covering transaction, behavior and complaint information. Core features including 30-day access frequency, consumption interval and consumption amount volatility are extracted. A Random Forest model is utilized to construct user churn prediction rules: the feature importance ranking is generated through the multi-decision tree voting mechanism, with an example feature split threshold being "consumption amount < 100 yuan and no access behavior for 15 days". Meanwhile, the Area Under the Receiver Operating Characteristic Curve (AUC, which evaluates the overall ranking capability) and Recall Rate (Recall Rate = correctly identified churned users / actual churned users) are calculated synchronously. After cross-validation, the model meets the criteria of AUC ≥ 0.85 and Recall Rate > 80%. It outputs churn warning lists and core rules (e.g., no access in the past 30 days, consumption amount decreased by 50% or more, and no after-sales interaction) to support operational actions such as targeted coupon distribution and product recommendation strategy adjustment.

搜集汇总

背景与挑战

背景概述

该数据集为电商用户流失预测训练模型，规模2G，周更新，基于随机森林算法预测用户流失，适用于电商平台用户行为分析。