Livestream e-commerce has been developing at a tremendous pace in recent years. On livestream platforms, such as Douyin, a retailer attracts viewers into the live room through short video advertising, and then streamers promote and sell products in real time. In such a scenario, an accurate prediction of traffic and sales plays an essential role in operation management, including live content strategy and inventory control. However, complex behaviors (follow, share, comment, etc.) of users and long conversion paths (from seeing the advertisement to entering the live room, and to purchasing the goods) lead to poor performance of traditional prediction methods. Additionally, few studies focus on advertising information in evaluating live room performance. Therefore, we propose a two-stage learning model for traffic and sales prediction based on behavior mining, which combines marketing models and deep learning methods. In the first stage, we integrate user behaviors before getting into the live room with short video advertising data for traffic prediction. In the second stage, based on the traditional marketing model, AIDA (Attention-Interest-Desire-Action), we design a funnel convolutional neural network (FCNN) to learn sophisticated behaviors in the live room in both time and behavior orientations, and take the predicted traffic volume as the auxiliary information for sales prediction. Extensive experiments on real-world datasets from Douyin illustrate the efficacy of our proposed method, which shows the value of fusing marketing models with deep learning techniques. Furthermore, the in-depth analysis provides practical insights into user behaviors for livestream e-commerce.

近年来，直播电商发展迅猛。在抖音等直播平台上，零售商通过短视频广告吸引观众进入直播间，然后主播实时宣传和销售产品。在这种场景下，流量和销量的准确预测对于运营管理（包括直播内容策略和库存控制）起着至关重要的作用。然而，用户行为复杂（关注、分享、评论等），转化路径长（从看到广告到进入直播间，再到购买商品），导致传统预测方法的性能较差。此外，很少有研究关注广告信息来评估直播间的表现。所以，我们提出了一种基于行为挖掘的流量和销售预测的两阶段学习模型，该模型结合了营销模型和深度学习方法。第一阶段，我们将进入直播间前的用户行为与短视频广告数据进行整合交通预测。第二阶段，基于传统营销模型AIDA（Attention-Interest-Desire-Action），我们设计了一个漏斗卷积神经网络（FCNN）来学习直播间中时间和行为方向上的复杂行为，并采取预测的客流量作为销售预测的辅助信息。对抖音真实世界数据集的大量实验说明了我们提出的方法的有效性，这显示了将营销模型与深度学习技术融合的价值。此外，深入的分析为直播电商的用户行为提供了实用的见解。