MARS: A Multi-task Ranking Model for Recommending Micro-videos

Jiageng Song, Beihong Jin, Yisong Yu, Beibei Li, Xinzhou Dong, Wei Zhuo, Shuo Zhou

Abstract: Micro-videos have become very popular recently. While using a micro-video app, the user experiences are strongly affected by the ranking of micro-videos. Moreover, the micro-video recommendation is often required to satisfy multiple business indicators. The existing models mainly utilize multi-modal features whose acquisition cost is too high for start-up companies. In the paper, we propose a multi-task ranking model MARS for recommending micro-videos. MARS aims at two tasks: finishing playing prediction and playback time prediction. For providing high accuracy in performing these two tasks, MARS adopts the multi-expert structure and mines historical statistical information besides interactions between users and micro-videos. Results of offline experiments and online A/B tests show that MARS can achieve good performances on two tasks. Further, MARS has been deployed in a real-world production environment, serving thousands of users.