Abstract
Fastest route recommendation (FRR) is an important task in urban computing. Despite some efforts are made to integrate A∗ algorithm with neural networks to learn cost functions by a data driven approach, they suffer from inaccuracy of travel time estimation and admissibility of model, resulting sub-optimal results accordingly. In this paper, we propose an ASNN-FRR model that contains two powerful predictors for g(⋅) and h(⋅) functions of A* algorithm respectively. Specifically, an adaptive graph convolutional recurrent network is used to accurately estimate the travel time of the observed path in g(⋅). Toward h(⋅), the model adopts a multi-task representation learning method to support origin-destination (OD) based travel time estimation, which can achieve high accuracy without the actual path information. Besides, we further consider the admissibility of A* algorithm, and utilize a rational setting of the loss function for h(⋅) estimator, which is likely to return a lower bound value without overestimation. At last, the two predictors are fused into the A∗ algorithm in a seamlessly way to help us find the real-time fastest route. We conduct extensive experiments on two real-world large scale trip datasets. The proposed approach clearly outperforms state-of-the-art methods for FRR task.