From Single Advertiser Optimality to Platform‑Wide Win‑Win: Introducing PlatformBid

In computational advertising, real‑time bidding (RTB) lets advertisers set high‑level goals (e.g., CPA, budget) while an automatic bidding algorithm decides bids; existing research and benchmarks such as iPinYou and AuctionNet focus on a single DSP perspective, optimizing each advertiser in isolation.

Problem Motivation

Modern ad platforms integrate SSP, DSP and the auction engine, so advertisers compete inside a unified system. This shift creates three gaps: (1) benchmarks no longer evaluate platform‑wide revenue and stability; (2) they cannot assess the impact when all advertisers adopt a new algorithm; (3) they miss dynamic competition when multiple advertisers adjust strategies simultaneously.

PlatformBid Benchmark Design

PlatformBid re‑defines the auto‑bidding objective by explicitly adding a platform‑level constraint \(O_l\) (e.g., platform revenue not below a baseline). Bids follow a multi‑parameter linear form updated every 30 minutes, leaving room for complex algorithms. Three evaluation settings model realistic business scenarios:

Homogeneous Competition : 48 advertisers use the same strategy, testing market‑wide equilibrium and potential efficiency loss.

Heterogeneous Competition : 24 baseline advertisers (vanilla Decision Transformer) vs. 24 test advertisers, mirroring gray‑scale A/B tests.

Promotional Competition : 7 advertisers double their budgets during a promotion (e.g., Black Friday), evaluating budget‑scale handling and CPA compliance.

Metrics are two‑fold: platform‑level (Conversion, Budget Utilization) and advertiser‑level (CPA Ratio, CPA Ratio Variance, Exceed Rate, Qualified Rate). A composite score Sc balances platform revenue with advertiser satisfaction.

Data and Implementation

PlatformBid builds on the AuctionNet dataset, extending it to 48 advertisers, 480 K training trajectories, and dense/sparse reward variants. A floor‑price mechanism (80 % of historical lowest price) prevents collusive under‑bidding.

BidFlow: Flow‑Matching Auto‑Bidding

Existing methods struggle with the multimodal bid distribution induced by platform‑level competition. Classic control (PID) cannot adapt; RL methods assume unimodal Gaussian policies; generative approaches are either deterministic or too slow for real‑time bidding. BidFlow addresses these limits by:

Using Flow Matching to learn a continuous transformation from Gaussian noise to bid actions, capturing complex multimodal distributions.

Guiding the learned policy with Q‑value estimates, enabling a single‑step distilled policy that outperforms the behavior‑cloned policy.

The architecture comprises a Critic network, a BC‑Flow policy trained with standard temporal‑difference learning, and a distilled single‑step policy optimized by (i) BC‑Flow distillation loss and (ii) Q‑value maximization. During inference, Euler integration solves the ODE in a few steps, yielding fast, one‑shot bids.

Algorithm Workflow

Training iterates over offline data: (1) estimate state‑action values, (2) train the flow‑matching network, (3) generate distilled targets via Euler steps, (4) update the single‑step policy, and (5) softly update target networks. After convergence, only the single‑step policy is deployed for online bidding.

Experimental Results

Setting 1 – Homogeneous Competition : On the Dense set, BidFlow scores 348.04, surpassing the next best baseline (GAS, 314.27) while achieving the lowest CPA exceed rate (0.25) and CPA variance (0.03). On Sparse data, it leads with a 30.59 score.

Setting 2 – Heterogeneous Competition : BidFlow attains the highest platform Conversion (347.96) on Dense data without harming baseline advertisers. On Sparse data, performance drops (Score 35.97 vs. DT 43.91) due to shared Transformer backbones causing low‑bid collusion, not an algorithmic flaw.

Setting 3 – Promotional Competition : BidFlow obtains the best scores for promotional advertisers (Dense 701.48) and non‑promotional advertisers (Dense 297.25), yielding the top overall platform score (356.20) while respecting CPA constraints.

Cross‑dataset validation on iPinYou confirms BidFlow’s superiority across all three settings, demonstrating dataset‑agnostic robustness.

Online Deployment

BidFlow was rolled out in Kuaishou’s e‑commerce ad system and compared against a heavily tuned Decision‑Transformer baseline. The live experiment showed a 0.68 % lift in expected consumption, confirming that the offline benchmark reliably predicts real‑world gains. The algorithm is now fully deployed.

Significance and Outlook

PlatformBid shifts auto‑bidding evaluation from “single‑advertiser optimal” to “platform‑wide win‑win,” providing the first platform‑level benchmark with realistic competition settings. BidFlow leverages Flow Matching and Q‑guided distillation to model multimodal bid distributions efficiently, achieving SOTA performance and demonstrating offline‑online consistency. The study also reveals that CPA‑ratio variance serves as a proxy for competition intensity, offering theoretical guidance for selecting cooperative‑type bidding strategies.