Neural Auction Mechanisms for Partial Fulfillment in Edge Computing

In the edge computing market, application service providers typically purchase computing resources from edge service providers to serve end users efficiently. In order to motivate edge service providers to actively deploy computing services, researchers utilize auctions to design an effective incentive scheme. However, traditional auctions assume all-or-nothing demand fulfillment and ignore scenarios where partial fulfillment is feasible. To address this challenge, two models are proposed: PDNet, built on fully connected neural networks, and EANet, which combines exchangeable matrix layers with attention mechanisms to learn optimal auction. Unlike prevailing methods, our approaches support flexible allocation, enabling partial fulfillment rather than all-or-nothing outcomes. The auction design is formulated as a constrained learning problem, yielding scalable and incentive-compatible mechanisms through end-to-end training. Extensive experimental results demonstrate that both PDNet and EANet outperform traditional baselines. EANet also exhibits strong generalization capabilities in scenarios with invisible input sizes. This work provides a modular and extensible deep learning framework for designing practical high-performance auction mechanisms in resource-constrained markets.