A Unified Guaranteed Impression Allocation Framework for Online Display Advertising

Editor’s note: Following the “technology efficiency” concept, Tencent Advertising has been exploring better exposure allocation algorithms to achieve a win‑win situation for advertisers and the platform.

At the top‑tier data mining conference ICDM 2022, the Tencent Advertising PTP contract and programmatic team collaborated with the University of Science and Technology of China LINKE Lab to propose a novel Unified Guaranteed Impression Allocation Framework (UGA) that simultaneously models and optimizes contract ads and real‑time bidding ads, thereby solving traffic competition between the two and improving global revenue.

1. Background

Online display advertising remains one of the most influential commercial models on the Internet, with a global market size of $189.3 billion in 2021. The market consists of two major business types: contract ads (Guaranteed Delivery, GD) and effect ads (Real‑time Bidding, RTB). Contract ads guarantee a fixed number of impressions for a predetermined price, while RTB ads are sold via auction.

Existing exposure allocation models treat these two types separately, leading to revenue loss due to logical separation. To address this, the paper introduces a unified framework that jointly optimizes both ad types.

2. Challenges

Huge commercial logic differences: the model must balance advertiser rights and platform revenue.

Complex competition: interactions among contract ads, among bidding ads, and between contract and bidding ads.

Non‑convex optimization: the problem is a Quadratically Constrained Quadratic Program (QCQP), an NP‑hard non‑convex problem.

3. Problem Modeling

The objective is to maximize the combined revenue of contract and bidding ads. The contract revenue includes guaranteed playback revenue, over‑delivery penalties, shortage penalties, and contract effect value. The bidding revenue follows the Generalized Second‑Price (GSP) auction mechanism.

Constraints cover contract exposure and click limits, bidding budget and conversion‑cost limits, and auction‑related GSP constraints.

Analysis shows that although the QCQP is NP‑hard, the bid variable b ij is the only base variable; all other variables can be derived from it, reducing the effective variable count to O(M) where M is the number of ads.

4. Framework Design

The UGA framework consists of feature enhancement and a differentiable sorting network (DSN) to approximate the QCQP solution.

4.1 Feature Enhancement

Bipartite matching: match ads with impressions based on targeting and attributes, preserving relevant features.

Feature transformation: aggregate exposure attributes for each ad and attach them as additional features.

Multi‑Layer Perceptron (MLP): compress high‑dimensional ad features to the dimensionality required by the bid formula weight w.

4.2 DSN‑based Loss Functions

The bid for ad j on impression i is computed as b ij = w·x ij + q ij , where w comes from the feature transformation module and q represents quality scores such as pCTR, pCTR·pCVR, etc.

Ranking loss: listwise optimization to improve overall GMV.

Effect‑metric loss: optimize individual ad performance metrics.

5. Experiments

Offline experiments under different contract reservation ratios (10 %, 20 %, 40 %) compare UGA with baseline LP, HWM, and SHALE. UGA consistently yields higher platform revenue, especially for contract ads.

Online A/B tests further compare UGA with LP, HWM, SHALE, M‑PID, and the latest USCB reinforcement‑learning algorithm. UGA improves both platform and advertiser revenue.

Convergence experiments show UGA converges faster and to a slightly better solution than USCB.

Ablation studies (UGA‑V1 without feature enhancement, UGA‑V2 without DSN and rank loss) quantify the contribution of each component.

6. Conclusion

The paper presents the UGA framework that jointly optimizes contract and bidding ads by modeling auction mechanisms and GSP constraints, providing an end‑to‑end solution for the non‑convex QCQP problem. Extensive offline and online evaluations demonstrate that UGA substantially increases total revenue for both advertisers and the platform.