Research Article
On the Influence of Grid Cell Size on Taxi Demand Prediction
@INPROCEEDINGS{10.1007/978-3-031-28813-5_2, author={Merlin Korth and S\o{}ren Schleibaum and J\o{}rg P. M\'{y}ller and R\'{y}diger Ehlers}, title={On the Influence of Grid Cell Size on Taxi Demand Prediction}, proceedings={Smart Objects and Technologies for Social Goods. 8th EAI International Conference, GOODTECHS 2022, Aveiro, Portugal, November 16-18, 2022, Proceedings}, proceedings_a={GOODTECHS}, year={2023}, month={3}, keywords={Taxi demand prediction Grid cell size Deep learning}, doi={10.1007/978-3-031-28813-5_2} }- Merlin Korth
Sören Schleibaum
Jörg P. Müller
Rüdiger Ehlers
Year: 2023
On the Influence of Grid Cell Size on Taxi Demand Prediction
GOODTECHS
Springer
DOI: 10.1007/978-3-031-28813-5_2
Abstract
Accurate taxi demand prediction has the potential to increase customer satisfaction and hence the usage of ride-sharing by predicting the number of taxis needed at a certain place and time. When reviewing the related work on demand prediction, we observed that in taxi demand prediction different grid topologies – e.g. rectangular subdivisions of an area – and sizes are applied. However, it is not clear how and why the grid cells are configured the way they are and a systematic comparison of different topologies and sizes as regards their influence on urban demand prediction is lacking.
In this paper, we compare the influence of different grid cell sizes – 250 m, 500 m, and 1000 m – on the prediction accuracy of different types of deep learning-based taxi demand prediction approaches, such as convolutional neural networks, recurrent neural networks, and graph neural networks. Therefore, we select five deep learning-based approaches from related work and evaluate their performance on the New York City TLC taxi trip dataset and three different evaluation metrics. Our results show that approaches with a grid cell of size 1000 m and 500 m achieve a higher prediction accuracy. Furthermore, we propose to consider the grid cell size as a tunable parameter in demand prediction models.
