A Context-Integrated Transformer-Based Neural Network for Auction Design

中文版

研究的问题：
这篇论文探讨的核心问题是如何设计一个激励相容的拍卖机制，以最大化拍卖人的预期收益。尽管理论上在单件物品拍卖中已有较为成熟的解决方案，但在多件物品拍卖中，如何设计最优拍卖机制仍然是一个挑战。近年来，深度学习方法在这方面取得了显著进展，但现有研究通常只关注固定数量的竞拍者和物品，或者将拍卖限制为对称形式。本研究通过将竞拍者和物品的公共上下文信息纳入拍卖学习框架，克服了这些限制。

研究的对象：
研究对象是具有上下文信息的拍卖设计，即在拍卖中，每个竞拍者和物品都有相关的上下文信息。这些上下文信息能够在一定程度上描述不同的竞拍者和物品，使得拍卖更接近实际情况。例如，在电子商务广告中，有大量具有各种特征的竞拍者和物品（即广告位），每个拍卖都涉及不同数量的竞拍者和物品。

使用的技术：
论文提出了CITransNet，这是一个基于上下文集成的Transformer神经网络，用于最优拍卖设计。CITransNet结合了出价信息、竞拍者上下文和物品上下文，开发拍卖机制。它建立在Transformer架构上，能够捕捉拍卖中不同竞拍者和物品之间复杂的相互影响。CITransNet在出价和上下文上保持排列等变性，即竞拍者（或物品）在出价信息和竞拍者上下文（或物品上下文）中的任何排列都会导致拍卖结果的相同排列。此外，CITransNet的参数数量不依赖于拍卖规模（即竞拍者和物品的数量），使其有潜力泛化到具有不同竞拍者或物品的各种拍卖中。

English Version

Research Problem:
The central problem explored in this paper is how to design an incentive-compatible auction mechanism to maximize the auctioneer’s expected revenue. Although theoretical approaches have provided optimal solutions for single-item auctions, designing optimal mechanisms for multi-item auctions remains a challenge. Recently, deep learning methods have made significant progress in this area, but existing studies typically focus on a fixed set of bidders and items or restrict auctions to be symmetric. This study overcomes these limitations by integrating public contextual information of bidders and items into the auction learning framework.

Research Object:
The research object is auction design with contextual information, where each bidder and item in the auction is equipped with relevant contextual information. This contextual information can describe different bidders and items to some extent, making the auctions closer to practical scenarios. For instance, in e-commerce advertising, there are a large number of bidders and items (i.e., ad slots) with various features, and each auction involves a different number of bidders and items.

Used Technology:
The paper proposes CITransNet, a Context-Integrated Transformer-based neural network for optimal auction design. CITransNet integrates bidding information, bidder contexts, and item contexts to develop an auction mechanism. It is built upon the Transformer architecture, capturing the complex interplay among different bidders and items in an auction. CITransNet maintains permutation-equivariance over bids and contexts, meaning any permutation of bidders (or items) in the bidding profile and bidder contexts (or item contexts) will cause the same permutation of auction results. Moreover, the number of parameters in CITransNet does not depend on the scale of the auction (i.e., the number of bidders and items), giving CITransNet the potential to generalize to auctions with various bidders or items.