Prediction And Application Of Travel Mode Selection For Urban Residents Using The Light Gradient Boosting Machine Method

In the transportation planning,the mode selection of residents 'travel is a difficult point to research.In the traditional "four-stage method",the mode selection uses the aggregate model to divide the mode selection of residents,ignoring the individual's travel behavior,which is lack of rationality.Then the disaggregate models are based on personal mode selection is developed,however,the assumption of random terms is not satisfactory.How to identify and quantify the influential factors and accurately predict mode selection is a hot point,which is also a problem that researchers urgently need to solve.Based on the research of disaggregate model,this paper put forward an improved Gradient Boosted Decision Tree(GBDT)machine learning method——the Light Gradient Boosting Machine(LightGBM),which is used to establish the model to predict the mode selection for the first time,and takes common factors in disaggregate models as the input variables to training models.Then,the model is evaluated by cross-validation method.The evaluation indexes were the average probability of correct assessment(APCA)and mean square error(MSE),and compared with other common machine learning models.The results show that the APCA of LightGBM is generally higher than the most models,and the APCA of the GBDT model is roughly the same,but from the view of running time,which is more suitable to deal with big data under the mode selection of residents.In addition,the model used to analyze the relative importance of each factor.The results can obtain the relative importance of the influencing factors,and select important influencing factors by the model.The model can use the single-factor variable method to study the law of the change of the different factors of the sharing rate of the bus.The travel mode selection model using the LightGBM can helps transportation researchers to make the reasonable prediction of the residents' travel modes,and can efficiently process massive amounts of data.It is also possible to further grasp the degree of influence and the law of various factors,which can provide a scientific basis for the transportation planning and the formulation of traffic policies.