Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Optimal Traffic Allocation for Multi-Slot Sponsored Search: Balance of Efficiency and Fairness

Anastasiia Soboleva, Alexander Ledovsky, Yuriy Dorn, Egor Samosvat, Andrey Tikhanov, Fyodor Prazdnikov

TL;DR

The paper addresses unfair traffic distribution in sponsored search by introducing Optimal Traffic Allocation (OTA), a two-stage, non-auction framework that jointly optimizes efficiency and fairness via a Gini-based metric. Stage 1 solves a per-query convex optimization to compute a target impression distribution, while Stage 2 uses probabilistic ranking to realize this distribution in real time; the design relies on a quadratic program with a Frank-Wolfe solver, ensuring online efficiency. OTA demonstrates superior fairness compared with baseline auction-based pacing methods while maintaining competitive efficiency, and it can serve as both an online allocator and an offline fairness-efficiency benchmark. The work delivers practical gains for real-time ad allocation and offers a pathway to hybrid systems that integrate OTA with traditional auctions.

Abstract

The majority of online marketplaces offer promotion programs to sellers to acquire additional customers for their products. These programs typically allow sellers to allocate advertising budgets to promote their products, with higher budgets generally correlating to improve ad performance. Auction mechanisms with budget pacing are commonly employed to implement such ad systems. While auctions deliver satisfactory average effectiveness, ad performance under allocated budgets can be unfair in practice. To address this issue, we propose a novel ad allocation model that departs from traditional auction mechanics. Our approach focuses on solving a global optimization problem that balances traffic allocation while considering platform efficiency and fairness constraints. This study presents the following contributions. First, we introduce a fairness metric based on the Gini index. Second, we formulate the optimization problem incorporating efficiency and fairness objectives. Third, we offer an online algorithm to solve this optimization problem. Finally, we demonstrate that our approach achieves superior fairness compared to baseline auction-based algorithms without sacrificing efficiency. We contend that our proposed method can be effectively applied in real-time ad allocation scenarios and as an offline benchmark for evaluating the fairness-efficiency trade-off of existing auction-based systems.

Optimal Traffic Allocation for Multi-Slot Sponsored Search: Balance of Efficiency and Fairness

TL;DR

The paper addresses unfair traffic distribution in sponsored search by introducing Optimal Traffic Allocation (OTA), a two-stage, non-auction framework that jointly optimizes efficiency and fairness via a Gini-based metric. Stage 1 solves a per-query convex optimization to compute a target impression distribution, while Stage 2 uses probabilistic ranking to realize this distribution in real time; the design relies on a quadratic program with a Frank-Wolfe solver, ensuring online efficiency. OTA demonstrates superior fairness compared with baseline auction-based pacing methods while maintaining competitive efficiency, and it can serve as both an online allocator and an offline fairness-efficiency benchmark. The work delivers practical gains for real-time ad allocation and offers a pathway to hybrid systems that integrate OTA with traditional auctions.

Abstract

The majority of online marketplaces offer promotion programs to sellers to acquire additional customers for their products. These programs typically allow sellers to allocate advertising budgets to promote their products, with higher budgets generally correlating to improve ad performance. Auction mechanisms with budget pacing are commonly employed to implement such ad systems. While auctions deliver satisfactory average effectiveness, ad performance under allocated budgets can be unfair in practice. To address this issue, we propose a novel ad allocation model that departs from traditional auction mechanics. Our approach focuses on solving a global optimization problem that balances traffic allocation while considering platform efficiency and fairness constraints. This study presents the following contributions. First, we introduce a fairness metric based on the Gini index. Second, we formulate the optimization problem incorporating efficiency and fairness objectives. Third, we offer an online algorithm to solve this optimization problem. Finally, we demonstrate that our approach achieves superior fairness compared to baseline auction-based algorithms without sacrificing efficiency. We contend that our proposed method can be effectively applied in real-time ad allocation scenarios and as an offline benchmark for evaluating the fairness-efficiency trade-off of existing auction-based systems.

Paper Structure

This paper contains 14 sections, 39 equations, 2 figures, 1 table, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related work
  3. Problem Formulation
  4. Reformulation of the Problem
  5. Algorithms
  6. Optimal impression distribution stage.
  7. Probabilistic ranking stage.
  8. Experimental Results
  9. Experiments on Synthetic Data
  10. Experiments on Real Data
  11. Metrics
  12. Baselines
  13. Evaluation results
  14. Conclusion

Figures (2)

  • Figure 1: Convergence of $OTA$ to the theoretical distribution by $OTD-FW$
  • Figure 2: Values of Efficiency and Gini metrics for the following algorithms: OTA, RCPacing, DMD and CTR-ranking. OTA algorithm is provided for different $\lambda$. PRPacing and DMD are provided for different impression amounts within budgets. The efficiency metric is presented as a relative value, normalized to the performance of CTR-ranking.