随着网络直播行业不断发展，越来越多的人拿起智能手机，成为一名主播或直播的忠实观众，网络直播打赏行为也随之成为国内一种独特的内容消费模式。目前对用户的网络直播打赏行为的研究尚比较缺乏。本文对网络直播平台用户打赏行为进行分析，预测用户是否对主播进行打赏，有助于了解用户兴趣，增加用户粘性，提升用户留存，并为网络平台的主播推荐策略提供帮助。本文挖掘了多个用于打赏行为预测的用户行为、主播行为特征，并提出了同时使用用户历史行为序列和主播历史行为序列学习用户、主播偏好的深度学习模型，对用户打赏主播的行为做预测。不同于其他领域的物品点击预测模型，作为预测对象的主播的特征更加多变，因此本文引入了主播历史被打赏的用户序列作为特征输入，实验发现可以提升模型的预测精度。本文对多任务学习的模型结构进行探究，多任务学习的优势在于可以使用更丰富的用户交互行为信息，对稀疏的打赏行为兴趣信息进行补充。真实数据集上的实验表明，将用户点赞数量和用户是否聊天作为辅助任务可以提升预测效果。和直接输出多任务预测结果的模型结构相比，采用将辅助任务的输出作为主任务输入的模型结构，并同时停止主任务的梯度反向传播到辅助任务的独立参数部分，可以达到更优的打赏行为预测效果。本文对特征的重要性进行了分析，结果表明：用户每次观看平均打赏次数、用户对该主播平均每次直播打赏次数、用户对该主播总聊天条数在预测时重要性较高。重要特征主要反映了用户的打赏频率，用户对该主播的付费意愿和忠诚度。同时，用户观看多样性特征对用户打赏行为有负向影响。

With the development of online livestreaming industry, more and more people pick up their smart phone and become an anchor or a loyal audience. Gift-sending behavior then turn into a unique content consumption pattern in China. Now, there is still a lack of research on users' online live broadcast gifting behavior. This paper analyzes gift-sending behavior in online livestreaming platform, predict whether a user will buy gifts for an anchor, which helps learn user interest, increase user engagement, improve user retention, and provide help for anchor platforms’ anchor recommendation strategy.The paper mines a number of user and anchor history behavior features, and proposes a deep learning model that uses both user and anchor history behavior sequence to capture user and anchor preference for gift-sending behavior prediction. Anchors have more variable personal characteristic than items, so the paper introduce anchor behavior sequence to learn their characteristics. Experiment shows that the model can effectively improve the prediction accuracy. The structure of multitask learning is also tried, which uses richer user interaction information to supplement sparse gifting interest information. Experiments on real datasets shows that taking user like amount and whether the user chats as auxiliary tasks can improve prediction result. Compared with the model structure that directly outputs the multitask prediction, the structure that uses the output of the auxiliary task as the input of the main task, and stops the gradient back-propagation of the main task to the independent parameter of the auxiliary task, can achieve better performance. The paper analyzes the importance of features, and the results show that the average number of user gifting per viewing, the average number of user gifting per live to the anchor, and the total number of chats of the anchor are of high importance in prediction. The important features mainly reflect the user's gifting frequency, and the user's willingness to pay and loyalty to the anchor. At the same time, user viewing diversity has a negative impact on users' rewarding behavior.