Multi-Objective Optimization for Guaranteed Delivery in Video Service Platforms

This article presents a practical study on content flow management for new video releases, focusing on how to allocate limited exposure resources to maximize playback volume.

Business Background: Video platforms face a scarcity of exposure slots across various scenes (home page, channel page, etc.). Each video competes for these slots, and the overall resource pool is limited, making a fair and efficient allocation essential.

Exposure Sensitivity Model (P2C): By analyzing historical PV‑click‑CTR logs of newly popular videos, the authors construct a differential‑equation‑based model (pv‑click‑ctr, abbreviated P2C) that captures the non‑linear, chaotic relationship between exposure (PV) and clicks. The model incorporates a saturation effect where additional exposure yields diminishing click returns.

The ODE formulation leads to the following relationships (equations 1‑4 in the original text), where parameters such as the click saturation value and growth rate are estimated via least‑squares fitting after data preprocessing (filtering, scaling, and transformation of sample points).

Guarantee (Retention) Model & Algorithm: Building on the P2C model, a multi‑objective non‑linear optimization problem is defined under exposure‑resource constraints. The objectives are to maximize total video views (VV) across scenes while minimizing the variance of CTR to ensure exposure fairness. Because the problem is non‑convex, a genetic algorithm (GA) is employed, with the fitness function derived from the P2C predictions.

The overall framework consists of two stages: (1) fitting the P2C model for each video to obtain its PV‑click function, and (2) solving the multi‑objective allocation problem using GA. Figures in the original document illustrate the model architecture and optimization pipeline.

Experimental Results: Offline experiments show that the P2C model achieves lower RMSE and APE compared with a baseline smoothing‑CTR method. Online bucket tests (30‑day and 7‑week periods) demonstrate that the proposed strategy reduces CTR variance by approximately 50% and improves overall scene CTR relative to manual allocation strategies.

Conclusion & Outlook: The proposed guarantee strategy effectively balances limited exposure resources with high‑demand content, delivering higher value across scenes. Future work includes extending the approach to PUV guarantee and addressing cold‑start challenges.