Less Traffic, Better Outcomes: Competition-Aware Re... | AI Research

Real-time bidding (RTB) ad exchanges typically forward nearly all incoming requests to demand-side platforms (DSPs), even though most do not result in a bid. This practice creates inefficiencies, as DSPs often throttle participation due to compute and budget constraints. The paper "Less Traffic, Better Outcomes: Competition-Aware Request Dispatch in Real-Time Ad Exchanges" introduces a framework that uses predictive modeling and probabilistic forwarding to selectively dispatch requests to DSPs, aiming to improve auction quality rather than simply maximizing traffic volume.

Improving Auction Efficiency

The authors, representing researchers from Huawei and H-Partners, propose a system that models the value of bidder participation. Instead of forwarding all requests, the framework estimates the probability of receiving a bid and the expected value of that bid for each request–DSP pair. By combining these signals, the exchange can prioritize sending requests to DSPs that are more likely to contribute to a competitive auction outcome. This approach treats request dispatch as a curation mechanism that balances the quality of participation against the quantity of traffic.

Technical Framework

The system operates through three primary components:

• Distributional Bid Modeling: A deep learning model (Bias-LF-DCN) estimates the probability of a bid and the distribution of potential bid values. To maintain low latency, the system uses a fusion architecture that shares request features across DSPs, reducing CPU usage.

• Probabilistic Dispatch: The exchange calculates an opportunity value for each request and uses a per-DSP threshold to determine forwarding probability. This includes an exploration floor to ensure the system continues to gather data for future model updates.

• Adaptive Optimization: Because market conditions are non-stationary, the system uses Proximal Policy Optimization (PPO) to periodically update dispatch thresholds offline. This allows the framework to adapt to changes in DSP behavior, budget constraints, and competitive dynamics.

Production Results

The framework was tested on a production platform handling over 20 billion requests daily. In a full multi-DSP deployment, the policy reduced the volume of requests sent to DSPs by 34.2% while increasing net revenue by 4.6% over a 14-day period. The researchers noted that while aggregate metrics showed a slight decline in the highest bid, this was driven by the suppression of low-value traffic. Stratified analysis revealed that Mid-RPM traffic segments saw significant revenue gains, suggesting that the policy effectively surfaces the comparative advantages of different DSPs.

Considerations for Implementation

The authors emphasize that aggregate metrics can be misleading in RTB environments. Their analysis shows that the benefits of selective dispatch are highly heterogeneous across traffic segments. While the framework improved monetized outcomes and reduced infrastructure load, the authors note that the system relies on the assumption that DSPs will adapt their bidding behavior to the new traffic mix. Because the study did not causally isolate DSP strategy changes from traffic composition effects, the authors suggest that cross-DSP interaction patterns should be viewed as diagnostic. The framework maintains end-to-end dispatch latency below 7ms, making it suitable for high-scale, real-time environments.