融合多维交互特征的电商直播异常行为识别模型及风控策略研究

Research on an E-Commerce Live Streaming Anomaly Detection Model and Risk Control Strategy Incorporating Multi-Dimensional Interaction Features

摘要: 针对电商直播场景下恶意刷单、虚假互动和恶意套取平台优惠等异常行为影响平台交易秩序与用户体验的问题，本文构建了一种融合多维交互特征的电商直播异常行为识别模型。区别于传统主要依赖交易金额、消费频次和最近消费时间等静态交易特征的识别方法，本文结合直播场景的实时互动特征与账号关联特征，引入弹幕频率、停留时长、设备IP聚集度等变量，形成面向电商直播场景的改进特征体系。针对风控场景正负样本极度不均衡的问题，采用SMOTE算法进行过采样平衡处理，并系统对比了逻辑回归、随机森林与XGBoost模型在分类任务中的性能。实验表明，结合SMOTE处理的XGBoost模型异常识别效果最优，F1-score达0.976，AUC接近1.0。进一步通过SHAP可解释性分析发现，弹幕频率和设备IP聚集度对模型预测结果的贡献较高，说明在电商直播异常行为识别中，交互特征和网络关联特征相较于单一交易特征具有更强的判别作用。基于上述结果，本文进一步提出实时流计算监测、账号关系图谱关联分析和分级风控处置策略，为电商直播平台异常行为识别与风险控制提供参考。

Abstract: To address the issues of abnormal behaviors such as malicious click fraud, fake interactions, and malicious exploitation of platform discounts in e-commerce live streaming scenarios, which disrupt transaction order and user experience, this paper constructs an e-commerce live streaming abnormal behavior recognition model integrating multi-dimensional interaction features. Unlike traditional methods primarily relying on static transaction features like transaction amount, consumption frequency, and recent consumption time, this study combines real-time interaction features of live streaming scenarios with account association features, introducing variables such as danmu frequency, dwell time, and device IP clustering degree to form an improved feature system tailored for e-commerce live streaming. To tackle the extreme imbalance of positive and negative samples in risk control scenarios, the SMOTE algorithm is employed for oversampling and balancing. Systematic comparisons were conducted among logistic regression, random forest, and XGBoost models in classification tasks. Experiments demonstrate that the XGBoost model with SMOTE processing achieves the best abnormal recognition performance, with an F1-score of 0.976 and an AUC close to 1.0. Further SHAP interpretability analysis reveals that danmu frequency and device IP clustering degree contribute significantly to model prediction results, indicating that interaction features and network association features play a stronger discriminative role than single transaction features in e-commerce live streaming abnormal behavior recognition. Based on these findings, this paper further proposes real-time streaming computation monitoring, account relationship graph association analysis, and hierarchical risk control strategies, providing reference for abnormal behavior identification and risk management in e-commerce live streaming platforms.