Driver Behavior Modelling Based On Ridesourcing Big Data

As a representative of the sharing economy,ride-sourcing has produced a large amount of ride-sourcing data,which contains rich urban travel information.Ride-sourcing big data has great significance in many aspects like traffic problem discovery and management,trip characteristics mining,intelligent transportation and urban planning.With the emergence of different ride-sourcing platforms,drivers can serve on multiple ride-sourcing platforms simultaneously.This multi-homing behavior brings new challenges to modeling drivers’ decision-making behavior.Based on ridesourcing big data,this dissertation firstly excavates traffic state characteristics in urban road networks to obtain city-scale macroscopic traffic supply patterns in time and space.Then,the ridesourcing drivers’ multihoming behavior and sequential decision-making behavior are fully studied at the microscopic level.This research can provide theoretical basis and policy suggestions for balancing the supply and demand of the ride-sourcing market.This dissertation has completed the following research work:(a)This dissertation proposes an urban traffic state estimation framework based on ridesourcing GPS traces.High-resolution GPS traces are mapped onto the road network efficiently using a cell-based map-matching method.The Schatten p-norm matrix completion algorithm is employed to recover the estimated sparse spatio-temporal speed matrix at the network level.Then,congestion probability maps are drawn based on the recovered matrix.Results show that the Schatten p-norm matrix completion algorithm can yield satisfactory estimation results(6km/h in mean absolute error)even under a 90% missing data rate.Meanwhile,the congestion probability maps help identify congested links in the urban road network.(b)The static decision of ride-sourcing drivers’ multi-homing behavior is modeled based on the ride-sourcing order data.The multinomial logistic regression model parameters are calibrated by large-scale real-world ride-sourcing datasets.The drivers’ multi-homing behavior mechanism is explained through calibrated parameters,and the effects of factors on multihoming behavior are quantitatively analyzed.The results show that drivers’ multi-homing behavior is significantly affected by socio-demographic characteristics,income level,and working hours(such as driving age,average income per order,and average time gap per order).(c)The static decision-making model is extended to the dynamic decision-making model.Based on a high-order hidden Markov model(HO-HMM),a dynamic decision-making frameworkis developed for driver’s multi-homing behavior analysis,which uncovers the driver’s high-frequency decision mechanism.The parameters are calibrated based on the maximum likelihood recursive algorithm.The model simulates the individual driver’s decisionmaking process while considering the time-varying market conditions.The results have verified HO-HMM’s advantage in modeling the extended historical dependency of drivers’ decisions and uncovered the variations of drivers’ attitudes in different hidden states on different platforms.(d)Based on the Bayesian network(BN),a supply decision model was developed to explore ride-sourcing drivers’ supply behavior during the COVID-19 epidemic.The changes in ride-sourcing demand and supply between the epidemic and pre-epidemic periods were compared by statistical analysis.In addition,the Bayesian inference is employed to analyze the impacts of confirmed cases and closed management policies on drivers’ online decisions and online hours.The results show that the developed Bayesian network yields accuracy of over80% in predicting ride-sourcing drivers’ online decisions during the epidemic period.This method can be extended to analyze the supply decision behavior of ride-sourcing drivers in other emergencies.The theoretical contributions of this dissertation are threefold:(a)a novel urban traffic state estimation framework is developed based on ride-sourcing big data;(b)a dynamic decision-making model based on a high-order hidden Markov model(HO-HMM)is proposed for ride-sourcing drivers’ multi-homing behavior analysis;(c)a supply decision model on the basis of Bayesian network(BN)is desined to explore ride-sourcing drivers’ supply behavior during COVID-19 epidemic.The proposed methods significantlty promote the ride-sourcing data based reseach on traffic macroscopic system analysis and microscopic behavior modeling.In practice,the critical operational indicators for the ride-sourcing system are adequately defined and calculated,which is conducive for the regulatory platform to formulate the overall market regulation strategy.Meanwhile,it helps identify critical attributes and characteristics affecting driver supply decisions,provide a theoretical basis for coordinating supply and demand in the ride-sourcing market and ultimately improve travel service efficiency.