Intelligent Auction Mechanisms for Alibaba Display Advertising: AIDA Framework, Deep GSP, and Neural Auction

Guest Speaker: Zhang Zhiling, Senior Algorithm Expert, Alibaba

Editor: Xiong Pei, Central China Normal University

Platform: DataFunTalk

Introduction

In recent years, the internet advertising market has grown to a scale of hundreds of billions, with mechanism design playing a core role in ad placement. This article, based on Alibaba’s Alimama display advertising business, details the evolution of intelligent auction mechanisms.

1. Alimama Display Advertising Business Background

Alimama’s display advertising includes products such as Super Recommendation and Super Diamond Display, serving millions of advertisers and generating significant revenue for the group.

Ads are delivered across Alibaba’s internal media, external apps (e.g., Alipay, Youku), and third‑party platforms, enabling full‑chain precise marketing for advertisers and high‑quality ad experiences for consumers.

To meet diverse advertiser goals, Alimama offers contract‑based brand ads, CPM/CPC/CPA performance ads, and optimization products like OCPM, OCPC, BCB, MCB, as well as the unified budget solution OneBP.

2. Multi‑Objective Optimization in Large‑Scale Media Platforms

Advertisers pursue heterogeneous objectives (exposure, acquisition, conversion), while media aim to improve user experience and GMV. The auction mechanism must allocate impressions and set prices to satisfy these multi‑party goals while preserving economic properties such as incentive compatibility and equilibrium.

3. AIDA Intelligent Decision‑Allocation Framework

AIDA (Advertising Intelligent Decision‑Allocation) introduces a two‑layer collaborative algorithm: the lower layer is a bidding agent that monetizes traffic value for each advertiser’s multi‑objective; the upper layer uses an auction mechanism to optimize platform‑wide objectives, ensuring incentive‑compatible equilibrium.

The framework supports static auctions (GSP, VCG) and advanced deep‑learning‑based auctions such as Deep GSP and Neural Auction, and provides a platformized engineering stack for rapid iteration.

4. Deep GSP: Multi‑Objective Smart Auction

Inspired by Myerson’s theorem, Deep GSP enforces monotonic allocation and minimal payment while learning advertiser‑specific coefficients via a neural network. It achieves incentive compatibility for utility‑maximizing advertisers and Nash equilibrium for multiple participants.

Empirical results show that Deep GSP outperforms traditional GSP and uGSP on Pareto fronts, delivering double‑digit growth in information‑flow ad volume and significant CTR/ROI improvements during major campaigns.

5. Neural Auction: End‑to‑End Learnable Auction

Neural Auction replaces discrete ranking with a differentiable sorting engine, allowing gradients to flow from final performance metrics back to the ranking model. It consists of a Set Encoder, a Context‑Aware Rank Score function, and a Differentiable Sorting Engine.

Experiments demonstrate superior multi‑objective performance compared to GSP, uGSP, and Deep GSP, with notable ROI gains during peak promotional periods. The approach has been accepted at KDD 2021.

6. Platformization of AIDA

AIDA provides two key engineering modules: the graph‑based online service engine (AIDA agent) built on TensorFlow for high‑throughput serving, and an offline strategy solution integrated into Alibaba’s Star Cloud platform for stable, scalable deployment.

7. Future Roadmap

Deepen technical research to address more complex business scenarios.

Co‑evolve new use cases and emerging technologies (e.g., set auctions, multi‑goal pricing).

Optimize the advertising ecosystem, including cold‑start for new products and advertisers.

Explore foundational mechanism theory for multi‑objective, multi‑agent environments.

Open‑source successful strategies and publish academic papers.

In summary, the AIDA framework, together with Deep GSP and Neural Auction, demonstrates how deep learning and mechanism design can jointly advance large‑scale e‑commerce advertising, delivering both academic insights and tangible business value.