Forecasting Citywide Crowd Flows with Unbalanced Human Mobility Distributions

Abstract

Predicting the movement of crowd flows in the city remains an open research problem. This article proposes a framework to predict the crowd flows at the city macro-level, spatially based on unbalanced flow distributions. Compared to models in literature, our framework is simpler, less computationally heavy, and attains state-of-the-art prediction results. In our experiments, we selected four baseline models to demonstrate the effectiveness of our solution. By grouping various regions composing a city into clusters, our proposed framework decreases the error rate (measured by RMSE, Root Mean Squared Error score) by 34% from the best baseline model, for the first hour prediction. In addition, our solution demonstrates high flexibility in including other urban features such as holidays, weather and social events.