First Region Last Route Recommendation via Graph Attention Network for Taxis

ABSTRACT Taxi passenger‐seeking recommendation can effectively improve driver profit and reduce fuel consumption. Existing methods always directly recommend passenger‐seeking regions or passenger‐seeking routes. However, these methods still face a threefold challenge of data sparseness, local optimization, and low efficiency. Therefore, we propose a novel model of first region last route recommendation via graph attention network for taxis , named FRLR‐GAT . Firstly, an offline‐online (O2O) model based on extreme deep factorization machine (xDeepFM) is put forward for passenger‐seeking region recommendation because xDeepFM has strong sparseness resistance. A larger grid with the highest pick‐up probability is recommended offline to avoid falling into local optimization. In the larger one, a smaller grid is further provided online. Secondly, a graph attention network (GAT) based route recommendation model is presented to online provide passenger‐seeking routes in the smaller grid, so that the proposed model obtains higher efficiency. Experiments on multi‐source datasets show that compared with some existing state‐of‐the‐art methods, FRLR‐GAT is able to respectively reduce the mean cruising distance and the average cruising time by at least 19% and 22%, while the average recommendation time is only 92 ms.