서울시 공공자전거 운행 거리 예측을 위한 머신 러닝 모델의 비교 연구

Abstract

Cities in many countries offer public bicycle sharing services to help solve the health and traffic jams of citizens and to solve environmental problems caused by automobiles. Using the machine learning models, the public bike service in Seoul have been analyzed. To service the bike sharing efficiently, the prediction of trip distance or trip duration will be needed. The prediction of trip distances and durations has important roles to help the proper operations and improvement of the bike rental services. The trip distances are deeply related to the trip durations, and could be calculated accurately when including the positions of pickup and return places, environment information such as temperature, humidity, fine dust density, et., al. To build models, linear regression, Random Forest, XGBoost and deep learning techniques have been used. Random Forest and XGboost provides important features among features. Especially, XGBoost being interested in many data manipulating and anlysing areas shows improved accuracy and can utilized the GPU to boost speed. To apply the deep learning in analysis of structured data(tabula data), embedding of categorical features will be done and the remaining continuous features and embedded features put into the fully connected neural nets. The deep learning neural net model shows the best accuracy and then XGBoost model, Random Forest models followed.