Recently, live streaming services attract millions of users’ participation and billions of capital investment. In each prevailing live streaming platform, there are thousands of anchors who are broadcasting concurrently, which means it is necessary for the platform to make recommendation to improve user experience. In such platforms, viewers change their watching preference dynamically, and anchors adjust their live content meanwhile. While there are many studies about predicting user’s (i.e., viewer’s) preference in literature, few methods proposed in literature can be used to predict live content’s change. As the recommendation target is online anchor’s live streaming that will be broadcasted in the next moment, we believe the prediction of the live streaming content is necessary for accurate recommendation. Therefore, in this paper, we study how to combine the prediction of viewer’s watching behavior and live content change for recommendation. We define a multi-task learning problem and propose a deep learning-based recommendation model, where we design two novel attention modules to capture viewer’s watching preference, anchor’s broadcasting preference, and loyal viewer’s preference related to each anchor. Experiments conducted on real datasets demonstrate the effectiveness of our proposed model.

近期，直播服务吸引了数百万用户的参与和数十亿的资金投入。在每个主流的直播平台中，都有成千上万的主播同时直播，这意味着平台需要进行推荐以提高用户体验。在这样的平台上，观众动态地改变他们的观看偏好，同时主播调整他们的直播内容。虽然文献中有很多关于预测用户（即观众）偏好的研究，但文献中提出的方法很少可以用来预测直播内容的变化。由于推荐对象是下一刻即将播出的网络主播直播，我们认为准确推荐需要对直播内容进行预测。因此，在本文中，我们研究如何结合观众观看行为的预测和直播内容变化进行推荐。我们定义了一个多任务学习问题并提出了一个基于深度学习的推荐模型，我们设计了两个新颖的注意力模块来捕捉观众的观看偏好、主播的广播偏好以及与每个主播相关的忠实观众的偏好。在真实数据集上进行的实验证明了我们提出的模型的有效性。