Generative Auto-Bidding in Large-Scale Auctions via Diffusion Completer-Aligner

Abstract

Bid optimization strategy in auto-bidding is central to computational advertising, achieving notable commercial success by optimizing advertisers' bids within constraints. Recently, generative models have revolutionized auto-bidding by directly learning a policy from large-scale datasets. Among them, the diffuser is superior in tackling sparse-reward challenges, along with its trajectory stitching and explainability capabilities, making it well-suited for industrial auto-bidding. However, its performance could be limited by generation uncertainty, particularly regarding generations' dynamic illegitimacy and preference misalignment, which can lead to suboptimal bids and further cause poor performance when competing with other advertisers in highly competitive auctions. To address it, we propose a Causal auto-Bidding method based on a Diffusion completer-aligner framework, termed CBD. Firstly, we conduct a theoretical analysis and propose a completer to augment the training process with an extra random variable t representing the decision timestep for enhancing the dynamic legitimacy between adjacent states. Then, we employ a trajectory-level return model as an aligner to refine the generated trajectories in inference for better alignment with advertisers' objectives. Experiments across diverse settings demonstrate that our approach not only achieves superior performance on large-scale auto-bidding benchmarks, such as a 29.9% improvement of conversion value in the challenging sparse-reward setting, but also delivers significant improvements on an online advertising platform, including a 2.0% increase in target cost.