From Single-Advertiser Optimum to Platform-Wide Win‑Win: Introducing PlatformBid (KDD 2026)

Research Background: Platform Perspective

Real‑time bidding (RTB) is the core paradigm of computational advertising: when a user generates an impression opportunity, the ad exchange solicits bids from all competing advertisers, and the highest bid wins. Existing auto‑bidding research and benchmarks (e.g., iPinYou, AuctionNet) focus on a single advertiser’s conversion maximization from the DSP viewpoint.

In modern industrial systems, supply‑side platforms (SSP), demand‑side platforms (DSP), and the ad exchange are integrated, so advertisers compete inside a unified platform. This creates two unmet needs: (1) evaluating how platform‑wide revenue and advertiser fairness change when all advertisers adopt a new algorithm, and (2) capturing dynamic competition when multiple advertisers adjust strategies simultaneously.

PlatformBid Benchmark Design

Problem Definition : PlatformBid formalizes auto‑bidding from a platform perspective by adding an explicit platform‑level constraint O_l (e.g., platform revenue must not fall below a baseline). Bidding policies use a multi‑parameter linear form updated every coarse time interval (e.g., 30 minutes).

Three Evaluation Settings :

Homogeneous Competition : All 48 advertisers deploy the same policy and compete via GSP auction, reflecting a full‑rollout scenario.

Heterogeneous Competition : 24 baseline advertisers use a vanilla Decision Transformer, while 24 test advertisers use the evaluated algorithm, mimicking a gray‑scale test.

Promotional Competition : 7 advertisers double their budgets during a promotion (e.g., Black Friday) while all use the same algorithm, measuring budget‑utilization and CPA compliance for both promotional and non‑promotional advertisers.

Metrics : Platform‑level conversion and budget utilization; advertiser‑level CPA ratio, CPA ratio variance, exceed rate, qualified rate; and a composite Score that balances platform revenue with advertiser satisfaction.

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

BidFlow: Flow‑Matching Based Auto‑Bidding Method

Motivation : Existing methods struggle with the multimodal bid distribution induced by platform‑level competition. Classical control (PID) cannot adapt, reinforcement learning (BCQ, IQL) assumes unimodal actions, and generative approaches (Decision Transformer, diffusion) are either deterministic or too slow for real‑time bidding.

Architecture consists of three jointly trained components:

Critic Network : Trained with temporal‑difference learning to estimate state‑action values.

BC Flow Policy : Trains a behavior‑cloning flow‑matching model that maps Gaussian noise to actions via a learned velocity field; the action at time t is a_t = (1‑t)·a_0 + t·a.

Single‑Step Policy : Distills the multi‑step BC Flow into a single inference step and further maximizes the Q‑value, achieving efficient one‑step bidding.

The training loop initializes value networks, BC Flow, and the single‑step policy, samples offline trajectories, performs value estimation, flow‑matching updates, and uses Euler integration to generate distillation targets. After convergence, only the single‑step policy is deployed for online bidding.

Experimental Results and Analysis

Setting 1 – Homogeneous Competition : On the dense dataset, BidFlow attains a Score of 348.04, outperforming the next best baseline GAS (314.27), with the lowest CPA exceed rate (0.25 %) and CPA variance (0.03). Similar superiority is observed on the sparse dataset (Score = 30.59).

Setting 2 – Heterogeneous Competition : BidFlow achieves the highest platform conversion (347.96) on the dense set without harming baseline advertisers. On the sparse set, performance drops (Score = 35.97 vs. DT = 43.91) due to shared Transformer backbone causing low‑bid collusion, not a flaw in BidFlow.

Setting 3 – Promotional Competition : BidFlow leads both promotional advertisers (Score = 701.48) and non‑promotional advertisers (Score = 297.25), with the overall platform Score = 356.20, demonstrating robustness under extreme budget imbalance.

Key Findings :

Conversion improvements: BidFlow raises conversion by 95 % over PID in Setting 1 and triples it on the sparse set.

Variance as competition proxy: Lower CPA‑ratio variance correlates with higher composite Score, indicating “cooperative” strategies that benefit the whole platform.

Cross‑Dataset Validation (iPinYou) : BidFlow ranks first across all three settings, confirming that PlatformBid and BidFlow generalize beyond the AuctionNet source.

Online Production Validation

BidFlow was deployed in Kuaishou’s e‑commerce ad scene and compared against the strongest online baseline (tuned Decision Transformer with custom rewards). It delivered a 0.68 % increase in expected spend while preserving CPA constraints, proving offline‑online consistency. The method is now fully rolled out.

Significance and Outlook

The work redefines auto‑bidding evaluation from “single‑advertiser optimum” to “platform‑wide win‑win,” introduces a platform‑level benchmark, and leverages Flow Matching to build a high‑performance, low‑latency bidding algorithm. The observed link between CPA‑ratio variance and competition intensity offers a theoretical guide for future platform algorithm selection.