A Multi-Task Learning Algorithm for Non-personalized Recommendations

In this paper, we introduce a multi-task learning (MTL) algorithm for recommending non-personalized videos to watch next on industrial video sharing platforms. Personalized recommendations have been studied for decades, while researches on non-personalized solutions are very rare to be seen, which still remain a huge portion in industry. As an indispensable part in recommender system, non-personalized video recommender system also faces several real-world challenges, including maintaining high relevance between source item and target items, as well as achieving multiple competing ranking objectives. To solve these, we largely extended model-based collaborative filtering algorithm by adding related candidate generation stage, Two-tower DNN structure and a multi-task learning mechanism. Compared with typical baseline solutions, our proposed algorithm can capture both linear and non-linear relationships from user-item interactions, and live experiments demonstrate that it can significantly advance the state of the art on recommendation quality.