Contextual Generative Auction with Permutation-level Externalities for Online Advertising

This work investigates whether generative models can continuously improve online advertising auction mechanisms, moving from the traditional "estimate‑then‑allocate" paradigm to a generation‑based approach that directly optimizes the final allocation.

Traditional generalized second‑price (GSP) auctions rely on the click‑through‑rate (CTR) separability assumption and ignore externalities among ads displayed together. Recent deep‑learning‑based auctions (e.g., DNA, SW‑VCG) still follow the estimate‑then‑allocate design, which cannot capture permutation‑level externalities and thus fails to reach globally optimal solutions.

We first analyze the optimal auction under permutation externalities and prove that the Myerson optimal form still holds when allocation and payment rules are decoupled. Based on this insight, we propose the first generative auction framework—Contextual Generative Auction (CGA)—that models permutation externalities with a generator‑evaluator architecture.

The generator is an autoregressive model that sequentially generates ad placements, incorporating a permutation‑invariant set encoder and a permutation‑equivariant decoder (GRU). The evaluator predicts CTRs for the generated sequence, using a bidirectional LSTM and self‑attention to produce an externality‑calibration vector. A PaymentNet module learns the optimal payment rule by minimizing ex‑post regret, ensuring incentive compatibility (IC) and individual rationality (IR).

Training decouples allocation learning from payment learning: the evaluator is first trained with cross‑entropy loss on observed clicks, then frozen to provide reward signals for the generator via a policy‑gradient objective that combines self‑reward (virtual welfare) and externality‑reward. The payment module is subsequently optimized with an augmented Lagrangian to enforce IC constraints.

Extensive offline experiments on a large‑scale Taobao display‑ad dataset (500k training, 100k test logs) and online A/B tests demonstrate that CGA improves platform revenue per mille (RPM) by up to 3.2%, CTR by 1.4%, GMV by 6.4%, and advertiser ROI by 3.5% with only a 1.6% increase in inference latency, closely approaching the theoretical optimal auction.

These results confirm that generative models can effectively capture permutation‑level externalities and provide a practical, high‑performance solution for online ad auction design.