💻 computer science

Beyond Single Slot: Joint Optimization for Multi-Slot Guaranteed Display Advertising

This paper proposes a novel joint optimization framework for multi-slot guaranteed display advertising that addresses slot redundancy and contract imbalance through an offline bipartite matching approach with a contract roulette mechanism, demonstrating significant improvements in revenue and contract fulfillment via large-scale online A/B tests on the Meituan platform.

Original authors: Zhaoqi Zhang, Jiaming Deng, Miao Xie, Linyou Cai, Qianlong Xie, Xingxing Wang, Siqiang Luo, Gao Cong

Published 2026-05-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhaoqi Zhang, Jiaming Deng, Miao Xie, Linyou Cai, Qianlong Xie, Xingxing Wang, Siqiang Luo, Gao Cong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a popular food delivery app (like Meituan) as a busy restaurant menu. On this menu, there are several spots where ads can appear: a banner at the top, a box in the middle, and a list at the bottom.

The Old Way: The "Single-Seat" Mistake
Previously, the system treated each ad spot like a separate, isolated table. If a user opened the menu, the computer would look at the top spot, pick the best ad for it, then look at the middle spot, pick the best ad for that, and so on.

The problem? This approach was like a restaurant where the manager assigns the best waiter to Table 1, then the same best waiter to Table 2, and the same waiter to Table 3.

Redundancy: The user sees the exact same ad three times on one page. It's annoying and wastes space.
Bullying: The most popular, high-paying advertisers would hog all the best spots, leaving smaller advertisers with nothing.
Imbalance: Some parts of the menu got too many ads, while others got too few.

The New Solution: The "Grand Conductor"
This paper introduces a new system that acts like a Grand Conductor for the entire menu page. Instead of deciding on each spot one by one, the system looks at the entire page at once and makes a coordinated plan.

Here is how their two main tricks work:

1. The "Page View" Rule (The Traffic Cop)

The system sets a strict rule: "No matter how popular an ad is, it can only appear a certain number of times on this specific page."

Analogy: Imagine a concert hall with different seating sections. The conductor ensures that the VIP section doesn't get 50 rows of the same singer while the balcony gets none. They spread the singers out so every section gets a fair share of the performance.
Result: This stops the "rich" ads from taking over the whole page and forces the system to use the "boring" or empty spots for other advertisers, making the whole page more balanced.

2. The "Contract Roulette" (The Fair Spinner)

Sometimes, the system wants to show a specific ad, but that ad is eligible for multiple spots on the same page. The old system might have put it in all of them. The new system uses a "Roulette" wheel.

Analogy: Imagine a game show where a contestant wins a prize. If they are eligible to win three different prizes, the Roulette wheel spins once. It picks one prize for them to take home. They can't take all three.
Result: This ensures that on any single page, a specific brand appears only once. It prevents the user from seeing the same pizza ad three times in a row, which makes the experience much smoother.

The Results: A Win for Everyone

The team tested this new "Grand Conductor" system on Meituan's real platform with thousands of real users.

For the Shops (Merchants): Their ads were shown to the right people more often, leading to a massive jump in sales and return on investment (up to 42% better in some tests).
For the App (Platform): The app made more money per user (about 28% more) because the ads were distributed more efficiently.
For the Users: They saw fewer repetitive ads and a more diverse, interesting menu.

In a Nutshell:
The paper says that by stopping the computer from making decisions one-by-one and instead making a "group decision" for the whole page, they fixed the problem of repetitive ads and unfair spot-taking. This made the advertising system fairer, more profitable, and less annoying for everyone involved.

Technical Summary: Beyond Single Slot: Joint Optimization for Multi-Slot Guaranteed Display Advertising

Problem Statement

Guaranteed Display (GD) advertising is a critical mechanism for platform monetization. However, mainstream approaches predominantly rely on a single-slot assumption, where each ad contract is optimized independently for a specific slot to maximize metrics like Click-Through Rate (CTR) and Payment Conversion Rate (CVR). While effective in simpler environments, this paradigm is increasingly inadequate for modern platforms (e.g., Meituan, Taobao) where a single Page View (PV) contains multiple ad slots.

The paper identifies three critical limitations in existing multi-slot environments:

Lack of Coordination: Local, single-slot optimization leads to the overuse of popular slots and underuse of others, reducing overall delivery efficiency.
No Upper Limit on Impressions: Existing methods ensure minimum delivery but fail to cap maximum impressions, allowing high-priority contracts to monopolize premium positions and creating imbalanced allocations.
Redundant Exposures: Request-level exclusivity fails to prevent the same contract from appearing in multiple slots within a single page, causing duplicate impressions and degrading user experience.

Methodology

The authors propose a joint optimization framework that shifts from online greedy allocation to an offline bipartite matching problem at the page-view level. The framework consists of two primary modules:

1. Page View-Constrained Allocation Module
This module formulates the allocation between ad requests (supply) and contracts (demand) as a constrained optimization problem. The objective function includes:

Smoothness Regularization: Penalizes deviation between actual allocation and the normalized target delivery ratio to stabilize delivery.
Priority-Aware Reward: Encourages allocation to higher-priority contracts.
Interest-Aware Matching: Promotes assignments with higher estimated user-interest relevance.

Crucially, the authors introduce a novel Page View (PV) constraint ( $pv_i$ ) to limit the number of impressions each slot can serve. This prevents the over-exposure of high-traffic positions and enforces balanced traffic allocation. The problem is solved using a distributed online strategy that iteratively updates dual variables ( $\alpha, \beta, \delta$ ) based on Karush-Kuhn-Tucker (KKT) conditions, allowing for scalable, real-time adaptation without solving the computationally expensive primal problem directly.

2. Contract Roulette-Based Selection Module
To resolve one-to-many matching conflicts and ensure that each contract appears at most once per page, the system employs a selection mechanism comprising:

Slot-wise Retrieval: Uses a roulette sampling strategy based on delivery weights to recall candidate contracts, ensuring long-tail exposure and fairness.
Probabilistic Filtering: Scores candidates using a utility function (CTR $\times$ CVR). If a contract appears in multiple slots, only the instance with the best relative position is retained.
Adaptive Bidword Control: Dynamically selects bidwords for each slot. The first slot samples randomly, while subsequent slots preferentially reuse the previous bidword if valid ads exist; otherwise, they switch to an available alternative. This maintains contextual coherence while reducing redundancy.

Key Contributions

The paper outlines three main contributions:

Unified Joint Optimization Framework: A novel formulation of contract-slot allocation as an offline bipartite matching task at the page-view level, enabling globally coordinated fulfillment.
Industrial Deployment Modules:
- PV Constraints: A mechanism for fine-grained traffic balance across slots.
- Contract Roulette Mechanism: A selection strategy to handle assignment conflicts and eliminate redundant exposures.
Real-World Validation: Successful deployment and extensive online A/B testing within the Meituan advertising system.

Experimental Results

The framework was evaluated on the Meituan platform using two rounds of online gray-scale experiments (35% and 70% traffic splits). The results, measured via Difference-in-Differences (DID) analysis, indicate significant improvements over the production baseline:

Merchant Efficiency: DID analysis showed a 42.17% increase in Merchant ROI, a 29.13% increase in Payment ROI, a 7.67% increase in CTR, and a 23.35% increase in Payment CVR.
Platform Revenue: The Average Revenue Per User (ARPU) increased by 28.17% under the 70% traffic setting.
Contract Fulfillment: The Fulfillment Rate improved by 2.12%, demonstrating that the system maintains stable and reliable delivery even with expanded exposure and stricter constraints.

Significance

The paper claims that its framework addresses the fundamental limitations of single-slot modeling in modern, multi-slot advertising environments. By moving beyond isolated slot-wise decisions to a globally coordinated, page-view-level optimization, the approach successfully balances fairness, exposure diversity, and delivery stability. The results demonstrate that joint optimization not only enhances merchant returns and platform revenue efficiency but also improves the user experience by eliminating redundant ad exposures. The work provides a scalable foundation for industrial GD systems requiring fine-grained traffic control and robust contract fulfillment.