상세정보

Bike Sharing System(BSS) is now becoming a prominent solution for traffic congestion and air pollution in many metropolitan areas. In BSS, passengers can rent a bike at one station and return it at another. One issue of BSS is station congestion. Stations with excessive rental rates are difficult to borrow a bike while stations with high return rates have no vacant spots. One solution for this imbalance between supply and demand is bike repositioning where bikes are transferred from congested stations to starving stations. Accurate estimate of demands for each bike station is necessary for a robust reposition- ing algorithm. However, current studies regarding demand prediction don’t account for useful features such as periodicity, limiting their modeling capacity. In this paper, we propose a method based on graph convolutional networks that enjoys strong predictive power of newly introduced periodicity features. On ‘Ddareungii’ benchmark dataset, pro- posed approach achieves a state-of-the-art result with 25% lower test mean squared error compared to previous top performer. We further inspect effects of each feature by abla- tion study and our analysis shows that weather and stations’ location do not necessarily impact predictions in an autoregressive setting, contradicting with common intuitions.

1 Introduction 1
2 Backgrounds 3
 2.1 Graph Convolutional Network. 3
 2.2 Station-Level Demand Prediction. 4
 2.3 Grid-Level Demand Prediction  4
 2.4 Task Formulation 5
3 Methods 7
 3.1 Dataset Description  7
 3.2 Periodicity,Weather and Location  8
 3.3 Network Architecture  9
 3.4 Training and Evaluation 11
4 Analyses 13
 4.1 Ablation Study. 13
 4.2 Effects of Weather and Location in Non-autoregressive Setting 15
5 Conclusion 16
Bibliography 17