L2RR: Towards a More Diversified Navigation Service with Learning to Rank Routes Framework.

Vehicle routing under the road network constraint is the cornerstone of navigation services. Although some commercial navigation systems can provide users with multiple route choices for a given origin and destination (OD) pair, users still encounter two main problems frequently in real cases: 1) The candidate set is not adequately diversified, and the route quality is only evaluated by its distance/time/turns while the driving environment along is overlooked; 2) The system recommends multiple routes to users simultaneously, which may interrupt and burden them while driving. In this paper, we aim to incorporate the driving environment to model the route quality comprehensively and rank all candidate routes accordingly, offering users a more diversified and efficient navigation service. However, it is difficult to effectively model the complicated driving environment, and there also lacks explicit guidance on route evaluation and ranking. To this end, we propose a novel Learning to Rank Routes (L2RR) framework for route representation and ranking based on multi-modal data. Specifically, we first utilize trajectories, street-view images, and attributes to model the driving environment of roads and routes. Then, we design a multi-task learning mechanism to evaluate route quality comprehensively. Experiments on a real-world dataset demonstrate that the proposed framework can achieve superior performance and outperforms state-of-the-art baselines.