Please use this identifier to cite or link to this item:
Author: Brimos, Petros
Title: Traffic forecasting with graph neural networks
Alternative Titles: Πρόβλεψη κυκλοφορίας δρόμων με χρήση νευρωνικών δικτύων γράφων
Date Issued: 2023
Department: Πρόγραμμα Μεταπτυχιακών Σπουδών Αναλυτική των Επιχειρήσεων και Επιστήμη των Δεδομένων
Supervisor: Kalampokis, Evangelos
Abstract: In recent years, there has been a growing interest in traffic forecasting using Graph Neural Networks (GNNs). These models manage to capture the complex topology in the non – Euclidean space by using Graph convolutions, while dynamic temporal dependencies are captured by recurrent networks. Moreover, the efficiency of such models mostly depends on the availability of cleansed and trustworthy mobility data. However, the access to such data is limited or restricted in most cases. Over the last decades, governments and the public sector publish their data, including dynamic traffic data, freely accessible by all citizens without any restrictions. These vast amounts of Open Government Data (OGD) have the potential to be exploited by data intelligence applications, such as Graph Neural Networks for traffic forecasting. To that end, this study focuses on a single case using open traffic data of the Greek OGD portal for predicting future traffic flows. More precisely, data exploration and analysis demonstrated that the OGD dataset contains many missing values and anomalies for the first two years of operation. The quality of the traffic data improves significantly after August 2022 where the experiments of this thesis were conducted. Two GNN models are implemented for the selected time window namely: Temporal Graph Convolutional Network (TGCN) and Diffusion Convolutional Recurrent Neural Network (DCRNN). The two GNN – based models achieved approximately 50% decrease on all error metrics compared with other baseline models, demonstrating high prediction precision. The TGCN model specifically achieved the best performance in traffic flow prediction among all prediction horizons and all metrics compared with all other models. To the author’s knowledge this is the first time a study exploited OGD traffic data for traffic forecasting with Graph Neural Networks.
Keywords: Graph Neural Networks
Open Government Data
Traffic forecasting
Spatial – temporal prediction
Deep learning on graphs
Open traffic data
Deep learning
Information: Διπλωματική εργασία--Πανεπιστήμιο Μακεδονίας, Θεσσαλονίκη, 2023.
Rights: Attribution-NonCommercial-NoDerivatives 4.0 Διεθνές
Appears in Collections:ΠΜΣ Αναλυτική των Επιχειρήσεων και Επιστήμη των Δεδομένων (Μ)

Files in This Item:
File Description SizeFormat 
BrimosPetrosMsc2023.pdf3.15 MBAdobe PDFView/Open

This item is licensed under a Creative Commons License Creative Commons