Learning-Based Ad Auction Design with Externalities: Score-Weighted VCG Framework

Online advertising increasingly relies on auction mechanisms, yet most existing designs ignore externalities—how the presence of other items (including organic results) affects click‑through rates (CTR). This paper proposes a general learning‑based framework, Score‑Weighted VCG (SW‑VCG), that explicitly incorporates externalities.

The framework decomposes optimal auction design into two sub‑tasks: (1) learning a monotone scoring function from bid data, and (2) solving a weighted‑welfare maximization problem for allocation. Theoretical analysis shows that, under various externality‑aware CTR models, SW‑VCG satisfies incentive compatibility (IC) and individual rationality (IR) while achieving near‑optimal revenue and social welfare.

We instantiate the framework with a matching‑based allocation algorithm. The scoring function is learned via a neural network that maps advertiser features to a monotone score; the allocation problem reduces to a maximum‑weight bipartite matching, solvable in polynomial time (Hungarian algorithm). Pricing follows from the VCG principle, yielding exact IC prices.

Extensive experiments on synthetic data and a large‑scale Taobao real‑world dataset demonstrate that SW‑VCG attains >99% of the theoretical optimal revenue and consistently outperforms baseline mechanisms such as VCG, GSP, and AMA in both revenue and social welfare.

The study confirms that a data‑driven, externality‑aware auction design can substantially improve ad platform performance without sacrificing economic guarantees.