Recommendations for ride-hailing zones are crucial for matching drivers with passengers efficiently, improving mobility, and managing traffic effectively. However, current recommendation methods based on centralized machine learning suffer from privacy concerns due to the need for data sharing and aggregation. Furthermore, the lack of personalized results tailored to individual regions or drivers within a single architecture often leads to spatial supply imbalances in ride-hailing services. To address these challenges, we propose in this work Federated Geo-Aware Matrix Factorization (FedGeoLMF), a framework for vehicular-cloud collaborative federated learning that identifies vehicles with similar passenger patterns, aiming to enhance the accuracy and efficiency of recommendations by improving local nodes for federated learning. At local nodes, federated matrix factorization algorithms based on historical trajectory data of moving vehicles and fused geographical information to achieve local learning and recommendations are designed. Additionally, we introduce a communication mechanism tailored for federated learning models exchanged between servers and mobile vehicles, employing the Named Data Networking (NDN) protocol that streamlines parameter uploading and downloading processes, unveiling akin patterns within local federated matrix decomposition outcomes by harnessing the NDN routing algorithm, delivering personalized collaborative recommendations while upholding the confidentiality of user-sensitive data. Finally, we design an online recommendation method based on driver location and reachability matrices to recommend optimal ride-hailing zones for drivers in different locations. Experimental results demonstrate that the proposed method outperforms current baseline models in terms of accuracy and training efficiency and can efficiently provide personalized ride recommendations while ensuring privacy in the ride-hailing service domain.

中文翻译：

---

移动车辆云中的协同联合学习网约车乘客区推荐


网约车区域的建议对于有效匹配司机与乘客、改善流动性和有效管理交通至关重要。然而，当前基于集中式机器学习的推荐方法由于需要数据共享和聚合而存在隐私问题。此外，单一架构中缺乏针对各个地区或司机的个性化结果往往会导致网约车服务的空间供应不平衡。为了应对这些挑战，我们在这项工作中提出了联合地理感知矩阵分解（FedGeoLMF），这是一种车云协作联合学习框架，可识别具有相似乘客模式的车辆，旨在通过改进本地节点来提高推荐的准确性和效率用于联邦学习。在本地节点，设计了基于移动车辆历史轨迹数据和融合地理信息的联邦矩阵分解算法，以实现本地学习和推荐。此外，我们引入了一种为服务器和移动车辆之间交换的联邦学习模型量身定制的通信机制，采用命名数据网络（NDN）协议来简化参数上传和下载过程，通过利用 NDN 路由来揭示本地联邦矩阵分解结果中的类似模式算法，提供个性化的协作建议，同时维护用户敏感数据的机密性。最后，我们设计了一种基于驾驶员位置和可达性矩阵的在线推荐方法，为不同位置的驾驶员推荐最佳的乘车区域。 实验结果表明，该方法在准确性和训练效率方面优于当前的基线模型，可以有效提供个性化乘车推荐，同时确保网约车服务领域的隐私。