New Framework Automates Ad Creative Selection at Scale, Boosts Revenue 6%

Demand-side platforms (DSPs) in digital advertising face a tradeoff: generating more ad variations from a single source image can unlock additional bidding opportunities, but expanding creative supply indefinitely delivers diminishing returns on revenue. Researchers have now developed a deployed solution to automate this decision at platform scale.

How ACQ Works#

The new system, called Automated Creatives Quota (ACQ), operates in two stages. First, it forecasts how much revenue each photo will generate at different creative quota levels using a specialized multi-task prediction model. This model is built on an unbalanced binary tree architecture—a design choice that addresses the reality that revenue distribution across quota tiers is heavily skewed, with some quota levels generating far more value than others.

The second stage treats quota allocation as a constrained optimization problem. Rather than deciding quotas independently for each photo, ACQ formulates the challenge as a multiple-choice knapsack problem (MCKP), which allows it to allocate capacity efficiently across the entire platform while respecting global resource limits. The system then solves this optimization using a dual-based algorithm that runs at the scale required for real-time ad delivery.

Real-World Results#

The framework has been implemented and tested on Kuaishou's advertising delivery platform. Both offline experiments and live online trials demonstrated measurable impact: the system achieved a 6.20% increase in platform advertising revenue compared to existing approaches. This gain came from more intelligent allocation of creative quotas across millions of photos and advertisers.