Scholarlink

<p>Combinatorial auctions are widely used for resource allocation, where bidders often submit bids on bundles because of inter-item relations such as complementarity and substitutability. However, most mainstream methods overlook these inter-item economic relations. In this paper, we propose a <b>N</b>eural <b>C</b>ombinatorial <b>A</b>uction <b>M</b>echanism with <b>C</b>omplementarity and <b>S</b>ubstitutability (NCAM-CS), which systematically incorporates item dependencies induced by complementarity and substitutability into the auction model. Specifically, we design a relation-aware graph neural network to capture high-order economic attributes of items from both complementarity and substitutability networks. Moreover, we employ behavior-driven feature learning to obtain representations that reflect individual preferences for each bidder. Finally, we introduce a hierarchical attention network to efficiently model bidder–bundle interactions for allocation and payment. Extensive experiments demonstrate that our approach outperforms baselines in terms of revenue, while also improving the interpretability of auction decisions. This study provides a novel technical approach to auction mechanism design in complex economic environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

NCAM-CS: neural combinatorial auction mechanisms with item complementarity and substitutability

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月15日

查看全文
  • Yuanyuan Zhang,
  • Yonglong Zhang,
  • Junwu Zhu,
  • Mingxuan Liang,
  • Xueqing Li,
  • Mingwei Zhao

摘要

Combinatorial auctions are widely used for resource allocation, where bidders often submit bids on bundles because of inter-item relations such as complementarity and substitutability. However, most mainstream methods overlook these inter-item economic relations. In this paper, we propose a Neural Combinatorial Auction Mechanism with Complementarity and Substitutability (NCAM-CS), which systematically incorporates item dependencies induced by complementarity and substitutability into the auction model. Specifically, we design a relation-aware graph neural network to capture high-order economic attributes of items from both complementarity and substitutability networks. Moreover, we employ behavior-driven feature learning to obtain representations that reflect individual preferences for each bidder. Finally, we introduce a hierarchical attention network to efficiently model bidder–bundle interactions for allocation and payment. Extensive experiments demonstrate that our approach outperforms baselines in terms of revenue, while also improving the interpretability of auction decisions. This study provides a novel technical approach to auction mechanism design in complex economic environments.