Fed-STWave: A Privacy-Preserving Federated Taxi Demand Prediction Model

With the proliferation of big data and advancements in intelligent transportation, taxi services have emerged as one of the primary commuting modes, resulting in a substantial influx of data related to taxi orders. Numerous researchers have proposed algorithms that leverage historical order data to model future demand, with spatio-temporal graph neural networks showcasing exceptional performance in predicting taxi demand. However, the sensitivity and high value associated with taxi order data contribute to the formation of “data silos” among different enterprises, impeding model optimization through data sharing. Moreover, existing graph-based federated models lack robust privacy protection measures during the initialization of the global model.To address this challenge, this paper introduces Fed-STWave, a federated privacy-preserving STWave model. We employ a generalized graph construction algorithm based on latitude and longitude coordinates to develop a prediction model using STWave. Subsequently, through federated learning, the model facilitates participants in collectively training a global model without divulging local data. Participants then engage in local training based on the global model to enhance prediction accuracy. Furthermore, homomorphic encryption is applied to secure data during model initialization and federated training, ensuring the privacy and security of participant data and models. Experimental results demonstrate that the proposed algorithm typically reduces the root mean square error (RMSE) for each participant by 5%-10%, compared to traditional local data training. Thus, under the premise of safeguarding participant data privacy, the algorithm presented in this paper partially mitigates the challenges associated with “data silos”.