Regularized Categorical Embedding for Effective Demand Forecasting of Bike Sharing System

The value of sharing economy services is increasing every year, and demand forecasting based service operations are essential for sustainable growth. For effective demand forecasting, this study proposes a categorical embedding based neural network model. The performance of this model is better than the traditional one-hot encoding based prediction; however, there are difficulties in creating a generalized prediction model due to the possibility of over-fitting of training data. Accordingly, it is possible to predict optimal demand by showing regularized performance applying techniques such as Batch Normalization, Dropout, and Cyclical Learning to the neural network. This methodology is applied to the Bike Sharing System to forecast bicycle rental demand by stations. In addition, in order to use the characteristics of global learning categories, uniform manifold approximation and projection (UMAP)-based dimensionality reduction technique is performed on the embeddings. The dimension-reduced embeddings are projected on the coordinate plane and used for K-means based cluster analysis, thereby providing an effective analysis result for demand patterns.