Interpretable Spatio-Temporal Graph Neural Networks for Real-Time Bike-Sharing Demand Forecasting and Resource Optimization

Abstract

Proliferation of urban bike share systems is a challenge as well as an opportunity to an efficient supply-demand area in real time in terms of sustainable transport. The design of interpretable spatio temporal graph neural networks (STGNNs) to predict bike sharing demand at high spatio temporal resolution is discussed in this work, which allows to optimize resources, i.e. dynamic rebalancing of bike availability and redistribution among the stations or zones. As a continuation of the more recent developments of spatio temporal deep learning and graph based demand modeling, we present a framework of the spatial relationships between stations, demand dynamics through time, and exogenous factors (e.g., weather, time of day). Importantly, we focus on interpretability: in addition to accurate prediction, we combine methods that can be used to understand what spatial, time-dependent and contextual variables are responsible in predictions. Our method proves to be highly predictive and provides human-competent explanations of surges or drops in demand based on the experimentation on real-world data of bike-sharing - therefore enabling real time decision support to operators.