| Abstract |
Free floating car sharing is a popular rental model for cars in shared use. In urban environments, it has become particularly attractive for users who make short trips or who make occasional use of the car. Since cars are not uniformly distributed across city areas, monitoring the number of cars available within restricted urban areas is crucial for both shaping service provision and improving the user experience. To address these issues, the application of machine learning techniques to analyze car mobility data has become more and more appealing. This paper focuses on forecasting the number of cars available in a restricted urban area in the short term (e.g., in the next 2 hours). It applies regression techniques to train multivariate models from heterogeneous data including the occupancy levels of the target and neighbor areas, weather and temporal information (e.g., season, holidays, daily time slots). To contextualize occupancy level predictions according to the target time and location, we generate models tailored to specific profiles of areas according to the prevalent category of Points-of-Interest in the area. Furthermore, to avoid bias due to presence of uncorrelated features we perform feature selection prior to regression model learning. As a case study, the prediction system is applied to data acquired from a real car sharing system. The results show promising system performance and leave room for insightful extensions. |
