A Stacking-Based Ensemble Model for Prediction of Metropolitan Bike Sharing Demand

Abstract

Due to the climate crisis and the improvement of public transportation networks, countries around the world are strongly advocating the low-carbon traveling mode. Shared bike as a new business model has a positive impact on the urban environment and transportation. The ability to estimate the hourly demand for bike sharing with high accuracy is essential for metropolis to offer stable bike rental services. Presently, data mining and predictive analysis technology can be utilized to realize the forecast of the hourly demand of shared bicycles. Data used in this article include the Seoul bike rented count dataset and weather information. This paper discusses various machine learning models for rental bike demand prediction, including Linear Regression, Ridge Regression, Lasso Regression, K-Nearest Neighbor, Random Forest, Decision Tree Regression, Support Vectors Machine, and Gradient Boosting Decision Tree. Different parameter tuning methods have been applied to improve the performance of basic predictive models. In addition, the redundant and irrelevant features have been removed to improve the performance of each basic model. After evaluating the individual basic predictors, several competent basic predictors are selected to compose a stacking-based ensemble model. Experimental results show that the stacking-based ensemble model outperforms the basic predictive models in all indicators.