Ride-Hailing Pick-Up Area Recommendation in a Vehicle-Cloud Collaborative Environment: A Feature-Aware Personalized Clustering Federated Learning Approach

Abstract Ride-hailing pick-up area recommendations based on GPS data can significantly enhance driver efficiency, improve passenger experience, and optimize urban traffic management. However, direct access to sensitive driver and passenger data, lack of data sharing between ride-hailing platforms, and other barriers severely undermine the reliability of current models and increase the difficulty of training and building these models. Additionally, the centralized mining of driver pick-up data can lead to spatial imbalances in ride-hailing platform order supplies, resulting in the recommendation of the same pick-up areas to different drivers in cases with similar spatiotemporal characteristics. This scenario not only exacerbates competition in certain areas but also impacts the overall service efficiency of the platform. To address these issues, this paper proposes a feature-aware personalized clustering federated learning model based on vehicle-cloud collaboration. This model achieves data sharing while protecting privacy, providing personalized pick-up area recommendations for drivers, and incorporating a feature-aware mechanism to prevent uneven order distribution. Specifically, the study introduces a clustering analysis algorithm based on Dynamic Time Warping (DTW) to accurately align and measure the similarity of drivers' order time series, classifying driver order-taking behavior patterns for more comprehensive behavior pattern analysis and improving model accuracy and reliability. Furthermore, we propose a deep matrix factorization algorithm, AFFM, that integrates driver features and geographical information. This algorithm deeply learns the interaction relationships between drivers and geographical features, modulates the attention weights of driver features, and constructs personalized models tailored to different driver pick-up characteristics, thus enhancing the precision and effectiveness of recommendations. Additionally, we present a federated learning algorithm, CFL\AFFM, based on vehicle-cloud collaboration. This algorithm leverages cloud computing resources to compensate for the limited computational resources of vehicles, allowing the AFFM algorithm to effectively utilize DTW clustering results. By sharing global models and clustering parameters through the cloud platform, broader collaborative effects are achieved, enhancing the model's generalization ability and stability while protecting driver and passenger privacy. Finally, we design a reachability matrix that incorporates the driver's current location and features. Based on the predictions of the CFL\AFFM algorithm, reliable pick-up areas are recommended, ensuring personalized and efficient ride-hailing services.