Dynamic Incentive Mechanism Design for Distributed Learning in 5G Private Networks: A Generalized Benders Decomposition Approach

In the 5G private network environment, to incentivize data providers to contribute data and computational resources for enhancing the generalization capability and model accuracy of distributed learning systems, this paper proposes a dynamic incentive mechanism based on Generalized Benders Decomposition (GBD). We model the interaction between the system administrator and data providers as a Stackelberg game, where the incentive strategy of the system administrator is optimized to achieve maximum system revenue while meeting the generalization requirements. To address the coupling in the leader-follower problem, we design an algorithm based on GBD, enabling the system administrator to dynamically guide the resource contributions of data providers through iterative incentive adjustments. Experimental results demonstrate that the proposed dynamic incentive mechanism outperforms baseline schemes in terms of generalization capability, system revenue, resource contribution level, and model training accuracy.