Διπλωματική εργασία--Πανεπιστήμιο Μακεδονίας, Θεσσαλονίκη, 2025.
(EL)
Made available in DSpace on 2025-02-17T15:21:24Z (GMT). No. of bitstreams: 1
Final Thesis Zisis Petkoglou 28-01-2025.pdf: 8122494 bytes, checksum: a93fcc3728326db95704908e04d5703b (MD5)
Previous issue date: 2025-01-28
(EN)
The accurate prediction of house prices is a critical challenge in real estate, directly influencing decisions by policymakers, investors, and homeowners. This thesis explores the integration of Linked Open Government Data (LOGD) with advanced Graph Neural Network (GNN) models to predict and explain property price variations across Scotland's data zones. The study models the spatial relationships inherent in real estate data as a graph, where nodes represent data zones characterized by socio-economic indicators, and edges capture their adjacency. Employing GraphSAGE, the research conducts binary node classification to determine whether property prices in a zone exceed the average price.
In addition to predictive modeling, the main purpose of this thesis is to implement and compare Explainable Artificial Intelligence (XAI) methods. A comparative evaluation of GNNExplainer, PGExplainer, DummyExplainer, Captum Integrated Gradients, and Captum Shapley Value Sampling assesses their efficacy in identifying the most influential features and graph structures driving predictions. Results indicate that the use of GNNs outperforms traditional machine learning models by leveraging the spatial dependencies in the data, while XAI techniques provide critical insights into the decision-making processes of these models.
The findings contribute to the growing body of literature on applying machine learning to open government data and highlight the role of explainability in deploying AI for socio-economic decision-making.
(EN)
Submitted by ΖΗΣΗΣ ΠΕΤΚΟΓΛΟΥ (
[email protected]) on 2025-02-15T10:45:50Z
No. of bitstreams: 1
Final Thesis Zisis Petkoglou 28-01-2025.pdf: 8122494 bytes, checksum: a93fcc3728326db95704908e04d5703b (MD5)
(EN)
Approved for entry into archive by Κυριακή Μπαλτά (
[email protected]) on 2025-02-17T15:21:24Z (GMT) No. of bitstreams: 1
Final Thesis Zisis Petkoglou 28-01-2025.pdf: 8122494 bytes, checksum: a93fcc3728326db95704908e04d5703b (MD5)
(EN)
University of Macedonia
