基于结构–关系互学多维协同注意力网络的抖音电商用户流失预测分析

Churn Prediction for Douyin E-Commerce Users Based on a Structure-Relation Mutual Learning Network with Multi-Dimensional Collaborative Attention

摘要: 短视频与直播电商的兴起使抖音等平台成为重要的在线消费入口，但高获客成本与用户易流失并存，急需有效的用户流失预测方法。针对传统模型难以同时刻画个体行为特征与用户间关系结构的问题，本文提出一种结构–关系互学多维协同注意力网络(SRML-MCA)，用于抖音电商用户流失风险预测。该模型由两个分支组成：结构分支(SFE)以年龄、收入、RFM指标、停留时长、浏览页数、Newsletter订阅等多维行为与属性特征为输入，学习用户个体表示；关系分支(RFE)基于兴趣标签、品类偏好和地域信息构建用户关系图，在图神经网络上捕捉相似用户的群体模式，并通过互学式多维协同注意力在两分支之间进行信息交互和权重重分配，得到兼具个体差异与关系结构的综合表示。本文选取阿里巴巴天池大赛抖音电商用户特征数据集，在缺乏长期行为日志的现实条件下，基于最近登录天数与购买频次构建标准化综合风险得分，并以其中位数为阈值划分高低流失风险，形成二分类标签。实验结果显示，在五折交叉验证下，SRML-MCA在Accuracy、Precision、Recall、F1和AUC等指标上均优于逻辑回归、随机森林、XGBoost、GCN和GAT等基线模型，AUC达到0.9992，F1达到0.9839，且标准差较小，表明该模型具有较强的判别能力与稳定性。研究表明，将结构特征与关系特征通过互学协同注意力进行深度融合，是提升抖音电商用户流失预测效果、支撑平台精细化运营的有效途径。

Abstract: The rise of short-video and live-streaming e-commerce has positioned platforms like TikTok (Douyin) as critical gateways for online consumption. However, these platforms face the dual challenges of high customer acquisition costs and high user churn rates, creating an urgent need for effective user churn prediction methods. To address the limitation of traditional models—which struggle to simultaneously capture individual behavioral characteristics and relational structures among users—this paper proposes a Structure-Relation Mutual Learning with Multi-dimensional Collaborative Attention network (SRML-MCA) for predicting user churn risk in Douyin e-commerce. The model consists of two branches: The Structure Feature Encoder (SFE) takes multi-dimensional user attributes and behavioral features—including age, income, RFM metrics, session duration, page views, and newsletter subscription—as input to learn individual-level representations. The Relation Feature Encoder (RFE) constructs a user relation graph based on interest tags, category preferences, and geographic information, leveraging graph neural networks to capture group-level behavioral patterns among similar users. A mutual-learning multi-dimensional collaborative attention mechanism enables dynamic interaction between the two branches, allowing adaptive reweighting of features and neighbors. This yields a unified user representation that integrates both individual distinctiveness and relational context. We evaluate our approach on the Douyin e-commerce user feature dataset from the Alibaba Tianchi Competition. Under realistic constraints—specifically, the absence of long-term behavioral logs—we define a standardized composite risk score based on days since last login and purchase frequency, using its median as the threshold to create a binary churn label (high vs. low risk). Experimental results under 5-fold cross-validation show that SRML-MCA consistently outperforms baseline models—including Logistic Regression, Random Forest, XGBoost, GCN, and GAT—across metrics such as Accuracy, Precision, Recall, F1, and AUC: it achieves an AUC of 0.9992, an F1-score of 0.9839, and exhibits low standard deviation, demonstrating both strong discriminative power and high stability. This study demonstrates that deeply integrating structural and relational features through a mutual-learning collaborative attention mechanism is an effective strategy for enhancing churn prediction accuracy in Douyin e-commerce, thereby supporting data-driven, fine-grained user retention strategies.