Game-theoretic incentives for federated learning in traffic prediction: Balancing resource allocation and prediction accuracy via Stackelberg contracts

Federated learning-based traffic flow prediction has attracted growing interest in the field. Federated Learning (FL) provides a novel solution for privacy-preserving distributed training. However, designing a fair and efficient incentive mechanism to encourage collaboration among diverse participants remains a key challenge. This paper proposes an incentive mechanism for FL based on contract theory and the Stackelberg game. More specifically, our proposed method quantifies and differentiates rewards for participant contributions through contract design while using the Stackelberg game to balance resource allocation and profit competition between the server and participants. Additionally, this paper integrates an efficient local prediction model, WL-Transformer (Weighted Layer Transformer), to enhance participants’ local data modeling capabilities, thereby improving the accuracy and adaptability of the global model in traffic flow prediction tasks. Finally, experiments on the License Plate Recognition (LPR) dataset from Changsha, China, demonstrate the effectiveness of the proposed incentive mechanism in achieving high-accuracy traffic flow prediction.