AIGB: Generative Auto‑Bidding via Diffusion Modeling

In 2023 Alibaba Mama introduced AIGB, a new paradigm that formulates the auto‑bidding problem as a generative sequence decision task, marking the first application of generative AI in large‑scale advertising auctions. The approach has been deployed on the Alibaba advertising platform and reported in a KDD 2024 paper.

To stimulate further research, a large‑scale auction auto‑bidding competition was launched, with a dedicated AIGB track that received NeurIPS endorsement; Alibaba Mama obtained the sole right to host a NeurIPS competition and will open‑source the AIGB benchmark (a standardized large‑scale simulated bidding system and dataset).

Background: online advertising relies on auction‑based bidding; auto‑bidding aims to maximize traffic value under budget, ROI, and other constraints. Traditional solutions use reinforcement learning (RL) but suffer from instability and difficulty handling long‑horizon sparse rewards.

Generative models (Transformer, Diffusion) can capture complex correlations in bidding data. AIGB treats auto‑bidding as a strategy‑generation problem, leveraging the strong distribution‑fitting ability of generative models to directly learn the relationship between historical trajectories and optimization targets, avoiding RL’s value‑function estimation errors.

DiffBid, a diffusion‑based AIGB implementation, models the entire bidding trajectory (budget, spend rate, ROI, etc.) as a diffusion process. During training it fits the distribution of historical trajectories; at inference it plans a trajectory that satisfies the given objective and refines it with an inverse‑dynamics controller.

Online experiments show DiffBid yields significant gains: +3.6 % GMV on a Max‑Return task and +5.0 % GMV on a Target‑ROAS task while maintaining ROI, and it reduces trajectory volatility and premature budget exhaustion.

DiffBid also offers better interpretability (planned vs. actual trajectories) and stronger multi‑objective compatibility, allowing post‑training adjustment of goals and incorporation of business‑logic metrics.

The framework is already commercialized at Alibaba Mama, powering the Double‑11 shopping festival, and its open‑source benchmark is expected to foster further innovations in auto‑bidding research.

The upcoming NeurIPS 2024 workshop will showcase competition results and invite the community to explore new AIGB solutions based on various generative models.