Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

A spatio-temporal graph neural network framework for predicting usage efficiency of e-scooter sharing services

Chelule, Kelvin LU (2023) In Student thesis series INES NGEM01 20231
Dept of Physical Geography and Ecosystem Science
Abstract
This thesis focuses on predicting the usage efficiency of e-scooters through the application of a Spatio-Temporal Graph Neural Network (STGNN). The predictions made by the STGNN will be compared with classical machine learning (ML) methods, including Random Forest (RF), Artificial Neural Network (ANN), and Linear Regression (LR). To facilitate the prediction of usage efficiency, the Time to Booking metric was utilized. PyTorch Geometric Temporal was employed to handle the spatio-temporal data modelling. Two STGNN models, namely Long Short-Term Memory R-GCN (LRGCN) and Attention Temporal Graph Convolutional Network (A3T-GCN), will be implemented and evaluated. Initially, usage efficiency will be measured based on the Time to Booking at the... (More)
This thesis focuses on predicting the usage efficiency of e-scooters through the application of a Spatio-Temporal Graph Neural Network (STGNN). The predictions made by the STGNN will be compared with classical machine learning (ML) methods, including Random Forest (RF), Artificial Neural Network (ANN), and Linear Regression (LR). To facilitate the prediction of usage efficiency, the Time to Booking metric was utilized. PyTorch Geometric Temporal was employed to handle the spatio-temporal data modelling. Two STGNN models, namely Long Short-Term Memory R-GCN (LRGCN) and Attention Temporal Graph Convolutional Network (A3T-GCN), will be implemented and evaluated. Initially, usage efficiency will be measured based on the Time to Booking at the DeSo level. Exploratory variables in both time and space will then be calculated as features for the prediction task. Subsequently, the STGNN models and ML models will utilize these calculated features to predict the usage efficiency. In order to determine the significance of variables in the RF and ANN ML models, permutation feature analysis will be conducted. The results of this analysis indicate that the start hour of the trip and the Point of Interest (POI) category "Lodging" are the most influential variables. This highlights the importance of incorporating temporal variables in prediction tasks and demonstrates the effectiveness of the STGNN approach in handling complex predictive tasks. The overall results indicate that the STGNN models exhibit superior performance compared to the ML methods in predicting the usage efficiency of e-scooters (Less)
Please use this url to cite or link to this publication:
author
Chelule, Kelvin LU
supervisor
organization
course
NGEM01 20231
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Physical Geography, Machine Learning, Deep Learning, Urban mobility, Usage efficiency, Geomatics
publication/series
Student thesis series INES
report number
628
language
English
id
9136857
date added to LUP
2023-09-08 11:03:37
date last changed
2023-09-08 11:03:37
@misc{9136857,
  abstract     = {{This thesis focuses on predicting the usage efficiency of e-scooters through the application of a Spatio-Temporal Graph Neural Network (STGNN). The predictions made by the STGNN will be compared with classical machine learning (ML) methods, including Random Forest (RF), Artificial Neural Network (ANN), and Linear Regression (LR). To facilitate the prediction of usage efficiency, the Time to Booking metric was utilized. PyTorch Geometric Temporal was employed to handle the spatio-temporal data modelling. Two STGNN models, namely Long Short-Term Memory R-GCN (LRGCN) and Attention Temporal Graph Convolutional Network (A3T-GCN), will be implemented and evaluated. Initially, usage efficiency will be measured based on the Time to Booking at the DeSo level. Exploratory variables in both time and space will then be calculated as features for the prediction task. Subsequently, the STGNN models and ML models will utilize these calculated features to predict the usage efficiency. In order to determine the significance of variables in the RF and ANN ML models, permutation feature analysis will be conducted. The results of this analysis indicate that the start hour of the trip and the Point of Interest (POI) category "Lodging" are the most influential variables. This highlights the importance of incorporating temporal variables in prediction tasks and demonstrates the effectiveness of the STGNN approach in handling complex predictive tasks. The overall results indicate that the STGNN models exhibit superior performance compared to the ML methods in predicting the usage efficiency of e-scooters}},
  author       = {{Chelule, Kelvin}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Student thesis series INES}},
  title        = {{A spatio-temporal graph neural network framework for predicting usage efficiency of e-scooter sharing services}},
  year         = {{2023}},
}