基于多任务与用户兴趣变化的短视频用户行为预测算法

Short Video User Behavior Prediction Algorithm Based on Multi-task and User Interest

为预测短视频用户行为(如:查看评论,点赞,点击头像,转发),考虑用户兴趣变化,将排序后的用户历史行为序列作为语料库引入Word2Vec训练得到词嵌入模型,学习用户的动态兴趣,有效捕获用户兴趣的变化。通过特征工程构建的统计特征与词嵌入模型构建的用户动态兴趣特征输入多任务模型,并提出一种新的评价指标来评估模型的预测精度。实验结果表明,相较于shared-bottom、Wide&Deep、DeepFM,提出的考虑用户兴趣变化的MMoE模型具有最优的预测精度。

The user behavior of short video(such as viewing comments,likes,clicking on avatars,and forwarding)is predicted by considering the change of user interests.In this paper,the sorted user historical behavior sequence is introduced into word2vec as a corpus to train the word embedding model,learn the dynamic interests of users,and effectively capture the changes in user interests.The statistical features constructed by feature engineering and the user dynamic interest features constructed by the word embedding model are input into the multi task learning with multi gate mixture of experts(MMOE),and a new evaluation index W-uAUC is proposed to evaluate the prediction accuracy of the model.The experimental results show that compared with shared bottom,wide&deep and deepfm,the proposed MMOE model considering the change of user interest has the best prediction accuracy.