Designing Effective Short‑Video Recommendation Systems: Goals, Multi‑Objective Modeling, and Long‑Term Value

According to the 2020 China Online Audio‑Visual Development Report, over 9 hundred million users watch short videos in China, with 8.2 billion daily active users spending nearly two hours each day; Baidu’s Haokan Video sees an average usage time of 70 minutes and more than 30 billion plays.

The platform hosts millions of creators, including over 100 k high‑quality ones, and produces original series such as “越来越好看”.

1. What problems does the recommendation system solve?

Three roles exist in a recommendation platform: users, creators, and advertisers. The system must improve the experience for users and creators.

User side: Deliver highly personalized, “千人千面” experiences by leveraging rich user profiles, contextual signals, and interest expressions.

Creator side: Ensure high‑quality content receives ample distribution, retain creators, and enable content meritocracy, while macro‑level ecosystem controls maintain a steady flow of quality creators.

2. Full picture of the video recommendation system

After creators upload content, it enters a unified ranking pipeline. When a user opens the app, the system recalls relevant items and follows the principle “Make Everything Happens”, then applies coarse‑ranking, fine‑ranking, and fusion stages to select the final feed.

3. Capturing user interest in the new auto‑play interaction

In the auto‑play flow, users either stay on a video or swipe away. “Swipe‑away” is treated as a negative signal ("hurt"), while long watch time or full playback is a positive signal ("satisfied"). The system uses three signals: hurt, duration, and completion.

Additional explicit signals such as likes, follows, and collections are also incorporated and organized into a four‑quadrant diagram for the ranking model.

4. Multi‑objective ranking

Modeling evolved from shared‑bottom DNN to MMOE and finally to population‑aware MMOE, training separate experts for low, medium, and high activity users to avoid dominance by high‑activity samples.

After predicting multiple objectives, simple polynomial fusion can be used, but it lacks adaptability. The current approach employs deepES for scenario‑aware personalized fusion, perturbing model parameters to generate candidate sets and selecting the best based on a designed reward that considers device type, refresh rate, and other context.

5. Long‑term value (LTV) oriented recommendation

Beyond immediate consumption, the system aims to predict future value of content. Past interactions serve as training data; current recommendations should extend user interest, and the future value of a video (LTV) is attributed back to the current item.

Designing LTV involves: (1) locating related content and defining a decay factor; (2) fitting a model to estimate long‑term consumption time and value.

6. Open questions on multi‑objective design

Are all objectives equally valuable (e.g., a like vs. the tenth like)? Is the current objective optimal for the whole user group, or should we consider collaborative value across users? Can we directly model user retention as an objective?

These questions invite further research and discussion.