Personalized Route Recommendation For Vacant Taxis: A Hybrid Learning Framework

With the rapid development of economy in modern cities,the travel demand of citizens has been increasing.As an important part of public transportation system in modern cities,taxis enormously facilitate the transportation travels of citizens.Due to the increasing complexity of urban traffic networks,it has become one of the hot issues in the field of urban transportation that how to recommend a proper cruising route for a vacant taxi.On the one hand,it can reduce the waiting time of passengers and improve the passenger travel experience by recommending proper cruising routes for vacant taxis.On the other hand,it can effectively reduce the energy(gasoline or electricity)consumption of taxis in the vacant states and improve the operation efficiency of taxis.By mining the historical trajectories of taxis,the potential spatial-temporal rules of their behaviors and driving preferences can be found,so as to recommend proper cruising routes for them.In the route recommendation of vacant taxis,traditional methods usually recommend vacant taxis moving to the areas with large taxi demand in the historical records or the nearest fixed passenger locations,which help taxis increase their profits.However,there are many deficiencies in the existing methods.For example,if taxis are only recommended to move towards the areas with large taxi demand,these route recommendation methods do not take into account the pick-up conflicts and the driving cost of taxis arriving at the pick-up locations.When many vacant taxis move to the same area,the transportation resources will be wasted.In addition,the taxi drivers tend to move along the roads or streets which are familiar with.The historical cruising routes has obvious personalized characteristics.Personalized route recommendation for taxi drivers are more reasonable,accurate and efficient,and the recommended routes tally with the cruising tendencies of taxi drivers.However,the traditional recommended routes of taxis do not consider the personalized cruising tendencies of taxi drivers.To this end,we propose a Hybrid Learning framework based Profit-Maximizing Personalized Route Recommendation Method for vacant taxis(HL-PPRRM)to obtain and recommend the profit-Maximizing personalized routes(PMPRs)to vacant taxis.PMPR comprehensively considers the cruising tendency of taxi drivers and the taxi demand of passengers.The hybrid learning framework in HL-PPRRM consists of the local learning on vacant taxis and the global learning on cloud server.Firstly,the historical cruising routes are locally learned by vacant taxis to predict their personalized routes,and the occupied records are globally learned by cloud server to predict the future taxi demand.Then,the future taxi demand for passengers in each region is released to vacant taxis,so as to adjust the personalized routes according to the future taxi demand to yield PMPRs for them.Extensive simulations and comparisons demonstrate the preferable performance of HL-PPRRM,i.e.,with the hybrid learning framework,the average displacement error of PMPRs is very small,the profits of taxis are significantly increased and the taxi driver’s personalized cruise tendencies are satisfied.Finally,based on the above research results,this thesis also implements a prototype system for recommending routes to vacant taxis.This system relies on the development of C/S architecture and Vue framework to realize the recommendation function of the profit-maximizing personalized routes.In addition,while providing a visual interactive interface,the user’s data information can be properly managed,which is convenient for the future queries and data exports for users.This thesis uses a hybrid learning framework to recommend cruising routes to taxis in the vacant states,which not only satisfies the personalized cruising tendencies of taxi drivers,but also greatly improves the business profits of taxis.However,from the perspective of network complexity and regional division,when the regional division is too small,and the number of urban areas is too large,the problem of long model training must be solved.Therefore,it is necessary to make a proper tradeoff between model performance and model complexity,and the training time should be shortened as much as possible even if the regional division is small.