Generative Auto-bidding via Diffusion Modeling (AIGB)

Last year we introduced the preliminary concept of AIGB (AI Generated Bidding). After a year of research, the full solution was deployed on Alibaba’s advertising platform and achieved significant gains in large‑scale budget AB tests. The work has been accepted at KDD 2024, and this article summarizes the method.

The online advertising market exceeded $626.8 B in 2023, and automatic bidding is a key driver of growth. Bidding decisions must be made sequentially over massive impression opportunities, forming a long‑sequence decision problem. Existing reinforcement‑learning (RL) approaches model the problem as a Markov Decision Process (MDP), but error accumulation limits performance on long horizons.

AIGB proposes a new paradigm: instead of RL, it treats auto‑bidding as a trajectory‑generation task using a conditional generative model. The model jointly learns the distribution of bidding trajectories and the optimization objective, eliminating cross‑step error propagation.

In practice, the method encodes the bidding problem as a conditional generation problem. Given historical non‑optimal trajectories and a set of constraints (budget, CPC, ROI, etc.), the model generates a future bidding trajectory that maximizes the specified objective. Both diffusion models and Transformers can serve as the backbone; the conditional diffusion process is illustrated in the paper.

The concrete algorithm, DiffBid, is trained by maximum‑likelihood estimation on historical trajectory data. A classifier‑free guidance technique enables multi‑constraint conditioning, allowing the model to satisfy several objectives simultaneously while preserving high reward.

Extensive offline experiments compare DiffBid with baselines such as USCB (a widely used RL bidding method), BCQ, CQL, and IQL. DiffBid consistently achieves the highest cumulative reward across budget settings and data sizes, and it better leverages exploratory data. Online A/B tests on Alibaba’s ad platform show that DiffBid improves buy count by 2.72 %, GMV by 4.2 %, ROI by 5.55 %, and ROI win‑rate by 20.04 % while keeping latency low.

The paper concludes that generative‑model‑driven auto‑bidding offers superior performance, interpretability, and ease of integration with expert knowledge. With billions of advertising trajectories available, the approach opens a path toward large‑scale marketing decision models and future research in AI‑augmented bidding.