Application Of Spatial Statistical Models And Methods In Shared Bike Scheduling

Since 2014,in order to solve the ‘last mile' traffic problem,bike sharing has become an important means of transportation.Recently,most of bicycle rental business of bike-sharing companies has reached a stage of expansion.The business goals usually focus on improving the user's experience and obtaining large-scale and high-quality user groups.Therefore,bike-sharing industry has generated the scenarios of predicting users' demand.However,when a bike-sharing company enters the stage of steady development and has stable user groups,it usually focuses on obtaining profits.Intuitively,profits can be earned by increasing the usage frequency.In order to measure the usage frequency,bike-sharing industry has proposed the ‘turnover rate' metric.The prediction of turnover rate has become an emerging application scenario of Statistics and data mining in Internet big data traveling.It can guide bike-sharing companies to adjust bicycle dispatch strategies and obtain more profits.Therefore,it is of great significance to solve the problem of predicting turnover rate.The main work of this paper is to predicting the turnover rate of different areas.At present,the bike-sharing industry usually uses traditional machine learning methods,which often ignore areas' space-time features.However,as a kind of panel data,traveling data could be used more effectively by considering the space-time dependencies of different areas.Therefore,the methods of Spatial Statistics can be considered in this research.This paper proposes several predicting models based on Spatial Statistics.The main contributions are:(1)The traditional machine learning methods are used to predict turnover rate.Based on the areas' macro and micro riding features,this paper uses linear model and XGBoost Regressor to predict each area's turnover rate.(2)Spatial Autoregression idea is introduced into traditional methods in part(1).This paper uses Moran's I metric to measure the spatial agglomeration of turnover rate.Then,this paper proposes SAR-Linear and SAR-XGB models based on spatial autoregressive model.By comparing the results of two models with traditional machine learning models,this paper shows that considering spatial autoregressive features is effective in this project.(3)Network Vector Autoregression idea is introduced into traditional methods in part(1).Referring to the Network Vector Autoregressive model,this paper uses cycling data to construct the network adjacency relationship of different areas.Hence,this paper proposes NAR-Linear and NAR-XGB models.By comparing results of two models with traditional machine learning models,this paper shows that considering the network autoregressive features is effective in this project.(4)Finally,from the perspective of features and model combination,two improvement methods are proposed in this paper.Firstly,SAR features in(2)and NAR features in(3)are simultaneously incorporated into XGBoost.Multi-spatial XGBoost is proposed in this part.Secondly,the combined SAR-NAR model is proposed.Compared with SAR-XGB and NAR-XGB,the R-squared of two models are further improved,reaching 58.07% and 58.65% respectively.Although some progress has been achieved in this paper,there is still some that could be further improved:(1)The data used in this paper are macro and micro data of a city in one week,which only has a small space-time span.If data with a larger space-time span could be obtained,the validity and stability of the model could be ensured.(2)This paper only discusses models based on linear model and XGBoost Regressor.This paper proposes SAR-Linear,SAR-XGB,NAR-Linear,NAR-XGB,multi-spatial XGBoost and SAR-NAR models as improvements.Other methods such as CNN and Time Series models are not discussed in this paper.The predicting results of these possible methods could be further compared in the future.(3)This paper predicts turnover rate of different areas by only one single model.However,areas usually have rich POI information in practice.Based on POI information,areas could be clustered for prediction in the future.This is reserved for subsequent research.