Intelligent Video Budget Pacing System for Online Video Platforms

Video pacing and promotion are critical distribution strategies for video platforms. Unlike ordinary recommendation algorithms, promotion must guarantee a certain exposure volume for designated videos while minimizing negative effects on the recommendation system.

For platforms such as iQIYI, new videos, fan‑generated clips, e‑commerce videos, and various series require guaranteed exposure to maximize playback volume.

The feed has limited exposure resources, leading to competition among content. Over‑aggressive pacing can exhaust a video's quota within minutes, preventing it from reaching more suitable users and reducing consumption metrics. Conversely, insufficient pacing leaves videos under‑exposed. The challenge is to achieve uniform distribution and keep the impact on overall consumption as low as possible.

To address this, an ecosystem‑wide intelligent promotion system was designed, based on an improved Budget Pacing algorithm borrowed from computational advertising. The system treats a video's target exposure as a budget and uses probabilistic throttling combined with fine‑ranking position boosting to control the pacing.

Problem analysis shows two goals: (1) uniform distribution by controlling the delivery speed using the target exposure, current exposure, and exposure curve; (2) improve video consumption metrics while keeping the system‑wide impact minimal. Probabilistic throttling is chosen over bid modification because each exposure is a fixed cost (one view), making bid adjustments infeasible.

The framework divides a day into 5‑minute time slots (288 slots) and derives the overall exposure trend from historical data. For each slot, a cumulative exposure curve is built, and a position‑exposure curve is estimated for fine‑ranking. The system computes the expected exposure for the upcoming slot, determines the maximum feasible ranking position, and adjusts the fine‑ranking score (σ) to meet the expected exposure while preventing over‑delivery. The probability of a promoted video participating in competition for a slot is also regulated to avoid quota overshoot.

Online results demonstrate that the system supports daily exposure at the hundred‑million level. Exposure completion rates rose from below 5% to 65‑70% on the new platform and 50‑60% on the baseline, a roughly 20‑fold increase. The impact on average watch time decreased from 3.5% to 2.5%, and the impact on average video impressions dropped from 2% to 0.15%. The system has helped promote high‑quality content such as "奇葩说" and various verticals, enhancing creator retention.

In summary, the Budget Pacing‑based intelligent promotion system significantly improves video pacing tasks. Future work will focus on order optimization of the promotion pipeline and the introduction of guaranteed‑quota mechanisms to ensure target exposure completion.

References: 1. Agarwal D, Ghosh S, Wei K, et al. Budget pacing for targeted online advertisements at LinkedIn. KDD 2014: 1613‑1619. 2. Xu J, Lee K, Li W, et al. Smart pacing for effective online ad campaign optimization. KDD 2015: 2217‑2226.