How to Optimize Ad Traffic Allocation with Front‑Load Curves and PID Control

Order Priority

Guaranteed‑volume ad orders are divided into at least ten priority levels based on order type, price, and advertiser. When traffic cannot satisfy all orders, higher‑priority orders must be fulfilled first, and priorities are dynamically adjusted by the commercial team.

Frontload (提前投放)

Although not a formal business requirement, front‑loading is widely used to mitigate traffic volatility in guaranteed‑volume advertising. By allowing extra delivery early, the system avoids waste when later traffic is scarce, while still keeping the overall delivery within an acceptable range.

Dynamic Delivery Constraints

The system enforces rules such as no more than one ad from the same industry in adjacent slots, a user seeing a brand’s ad no more than twice per week, and a maximum of three ads from the same order in a video stream. These constraints protect user experience but complicate bipartite‑graph construction for pre‑delivery planning.

Partial Guaranteed Orders

Low‑priority, low‑price orders may remain incomplete but must not be delivered too quickly or finish too early. Their goal is to improve traffic utilization by preferentially selecting higher‑price orders within allowed limits.

Ideal Frontload Curve

The ideal front‑load curve is derived from a differential equation. Its initial slope is 1+θ , where θ is a configurable factor, and the slope decreases over time, ending at (1,1) to ensure 100% delivery by the deadline.

The curve’s initial slope is 1+θ , allowing delivery faster than uniform at the start.

The slope gradually declines, reaching full delivery at the order’s end time.

The solution of the differential equation is shown in the following image:

In practice, traffic distribution is non‑uniform and orders may have blackout periods. The horizontal axis is reinterpreted as the proportion of cumulative traffic up to a given time, allowing the same curve to represent realistic scenarios.

PID‑Based Selection Coefficient

The selection coefficient adjusts based on the deviation between actual and ideal delivery curves. Using a PID controller, the proportional term, derivative term, and (optionally) integral term are combined, though the integral term is omitted for simplicity.

The resulting update rule for the selection coefficient (log‑transformed to keep it positive) is illustrated below.

Probability Allocation Process

Each ad starts with zero selection probability. Two rounds of allocation occur: first based on the selection coefficient, then on the upper coefficient. If the sum of selection coefficients exceeds 1, high‑priority orders receive their full share and lower‑priority orders may receive zero probability. If the sum of upper coefficients is below 1, all ads receive probability equal to their upper coefficient, leaving unused probability slots.

Partial guaranteed orders are allocated solely based on their upper coefficients, ordered by unit price.

Practical Examples

Three illustrative examples demonstrate single‑round selection‑coefficient allocation, two‑round allocation with both coefficients, and allocation for partial guaranteed orders.

Implementation in Disney Streaming

The traffic allocation solution is split into three parts: computing the ideal delivery curve, calculating the selection coefficient, and determining the selection probability. Additional mechanisms handle burst traffic by limiting per‑ad selection counts.

While differing from classic ad‑matching algorithms, the approach strives to balance theoretical rigor with real‑world constraints, illustrating the need to adapt classic algorithms for production environments.