摘要：

出租车载客推荐能够有效提高司机利润, 对于提升交通效率、改善城市出行体验以及推动智能交通的发展都具有重要意义。现有方法一般直接向司机进行载客区域或载客路线推荐, 没有考虑将这两者进行结合, 不仅面临数据稀疏性问题, 而且难以兼顾推荐准确性与实时性能。为此, 提出一种面向GPS数据的出租车载客路线层次化推荐模型, 其中采用了抗稀疏性的极深因子分解机(xDeepFM)、深度Q网络(DQN)强化学习算法以及层次化推荐策略。首先, 离线推荐高载客概率的大网格, 以减少在线计算量; 然后, 当出租车司机提出实时载客推荐需求时, 在离线推荐的大网格内进一步推荐高载客概率的小网格; 最后, 给司机规划一条到小网格的载客路线。在滴滴公司数据集上进行实验, 结果表明, 与现有的一些先进方法相比, 该方法可以使空载出租车司机的巡航时间至少减少36%, 巡航距离至少减少26%, 并且推荐时间仅需85 ms。

关键词: 载客路线推荐, 载客区域推荐, 层次化推荐, 极深因子分解机, 深度Q网络

Abstract:

Taxi pick-up recommendations can increase driver profits, improve traffic efficiency, enhance urban travel experiences, and advance intelligent transportation systems. Existing methods typically recommend either pick-up areas or pick-up routes to drivers, without combining both, resulting in data sparsity and challenges in balancing recommendation accuracy with real-time performance. This study proposes a hierarchical recommendation model for taxi pick-up routes using GPS data incorporating a sparsity-resistant extreme Deep Factorization Machine (xDeepFM), Deep Q Network (DQN) reinforcement learning algorithm, and a hierarchical recommendation strategy. The proposed method first recommends a high-probability pick-up area (large grid) offline to reduce online computational load. When a taxi driver requests a real-time pick-up recommendation, a smaller high-probability pick-up within the offline-recommended large grid is suggested. Finally, a pick-up route is planned for the driver. Experiments on the DiDi dataset demonstrate that, compared to existing state-of-the-art methods, the proposed approach can reduce idle taxi drivers' cruising time by at least 36% and cruising distance by at least 26%, and the recommendation time is only 85 ms.

Key words: pick-up route recommendation, pick-up area recommendation, hierarchical recommendation, extreme Deep Factorization Machine(xDeepFM), Deep Q Network(DQN)