How to Maximize Video Views with a Multi‑Objective Exposure Optimization Model

1 Business Background

Guarantee (or "保量") strategies are crucial for video content promotion because new videos need to increase their own exposure to maximize play counts, while overall exposure resources across scenes (homepage, channel pages, etc.) are limited, creating competition for allocation.

Figure 1: Series channel page and homepage.

2 Exposure Sensitivity Model

To model the relationship between exposure (PV) and clicks, an ordinary differential equation (ODE) based PV‑click‑CTR (P2C) model is constructed. The model captures the saturation effect where additional exposure yields diminishing click increments, expressed by equations (1)–(4). Parameters are fitted using least‑squares on filtered historical PV and click data.

Figure 2: P2C model overall architecture.

3 Guarantee Model & Algorithm

Based on the fitted P2C model, a multi‑objective nonlinear optimization problem is formulated under exposure‑resource constraints for each scene and drawer. The objectives are to maximize total view volume (VV) across scenes and minimize CTR variance to ensure fair exposure. Because gradient‑based solvers are unsuitable, a genetic algorithm (GA) is employed, using the P2C model to evaluate fitness.

Figure 3: Guarantee model overall framework.

4 Experimental Results

Offline tests on several hot videos show that the P2C model achieves lower RMSE and APE compared with a smoothed CTR baseline. Online bucket experiments compare the proposed strategy against a manual guarantee strategy; over 30 days and 7 weeks the new strategy improves overall CTR and reduces CTR variance by about 50% on average.

5 Conclusion & Outlook

The guarantee strategy addresses the mismatch between limited traffic resources and abundant video demand by providing optimized exposure recommendations, improving both total views and fairness. Future work includes extending the framework to PUV guarantee and addressing cold‑start scenarios.

This work was accepted at KDD2020: Hang Lei, Yin Zhao, and Longjun Cai. “Multi‑objective Optimization for Guaranteed Delivery in Video Service Platform.” Proceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2020.