Dockless Shared Bike Distribution Inference And Hotspots Detection In New Cities

Dockless shared bikes,which aim at providing a more flexible and convenient solution to the first-and-last mile connection,come into China and expand to other countries at a very impressing speed.Just taking last year as an example,dockless shared bikes had been deployed in more than 100 cities of China.When expanding bike sharing business into a new city,most start-ups always have to figure out how to cover the whole city with a suitable bike distribution so as to maximize the usage of bike resources.To assist dockless shared bike expansion progress and allow possible actions to be taken in advance,this paper issues the problems of dockless shared bike distribution inference and hotspots detection when deploying bikes into a new city.Both two problems are challenging due to the time-varying bike demand distributions and characteristics of no fixed parking stations.As there are no dockless bikes being deployed in the new city,we propose to learn insights from cities that are populated with dockless bikes.We analyze the spatial-temporal characteristics from available dockless shared bike distributions and exploit multi-source urban data to identify discriminative features that affect bike distributions.However,the non-negligible feature space differences among cities obstruct the knowledge transferring progress,resulting in unexpected inference performance degradation.Therefore,in this paper,we develop a novel inference model named GeoConv which combines Factor Analysis and Geographic Convolutional Neural Network to reduce the city domain difference and capture the effects of the neighboring areas.As for hotspots detection problem,we design an end-to-end model equipped with adversarial learning progress and ranking loss optimized function to learn deep domain-invariant representations of origin feature spaces.The extensive experiments on real-life datasets show that the proposed solution provides significantly more accurate inference and detection results compared with the complex competitive prediction methods.