Distributional serving improves watch time prediction in videos
Beyond a Scalar: Distributional Serving Interfaces for Watch-Time Prediction
Information Retrieval
Summary
Predicting how long people watch short videos helps video platforms show more engaging content. The authors found that existing methods usually give only a single expected watch time number, missing detailed information like the chance of watching the whole video or only parts of it. They created a new system that represents watch time as a distribution capturing multiple viewer behaviors and event timings. This richer information helps make better predictions and improves ranking of videos in several real datasets.
What this means in practice
- •For video platform engineers: Improve video ranking systems by incorporating distributional watch time predictions instead of single expected values.
- •For recommendation system developers: Build task-specific lightweight models using distribution summaries to predict various watch-related targets like completion or overplay probabilities.
- •For online advertising teams: Estimate richer engagement metrics from watch time distributions to better target ads on video feeds.$Commercial implications: Provides a more detailed watch time distribution enabling advanced ad targeting, improving ad placements and revenue.
Authors
Xuan Liu, Jingbin Qian, Zhanyu Liu, Hefeng Zhou
Abstract
Watch time is the primary engagement signal in short video feeds, and its prediction directly affects ranking and exposure. Existing methods improve watch time prediction by correcting duration bias or modeling richer distributions, but most expose only an expected or debiased watch time at serving time. Even when video duration is available to later models, the interface gives only one estimate of watch time and no probabilities for completion, overplay, or other regions relevant to downstream tasks. To address this limitation, we propose the Distributional Serving Interface (DSI), which has a distribution provider, a compact, low-dimensional summary, and lightweight readouts tailored to each task. The provider learns a joint distribution over four watch states derived from watch ratio and their event times; rules based on video duration remove incompatible combinations, while a restoration loss preserves accuracy in seconds. The summary reduces this distribution to a small set of event probabilities, time scales relative to duration, and uncertainty statistics. After training the provider, we fix its parameters and train value and ranking readouts that combine the summary with raw context. Across KuaiRec, KuaiRand-1K, and WeChat21, the complete DSI system achieves the lowest MAE on all three datasets, beating the strongest result among nine baselines by 1.9% to 8.5%, and achieves the best XAUC on two. It also leads retrieval metrics that account for video duration when complete systems are compared. With matched readouts held constant, the summary retains information relevant to each task beyond a predicted mean paired with video duration. Using the same lightweight linear heads for each new target, it also performs best on two new watch-time targets and improves a separately logged engagement target, while a randomly initialized provider does not reproduce this gain.