User: Guest  Login
Original title:
On the Convergence of Structure and Geometry in Graph Neural Networks
Translated title:
Zur Konvergenz von Struktur und Geometrie in Graph-neuronalen Netzen
Author:
Gasteiger, Johannes
Year:
2023
Document type:
Dissertation
Faculty/School:
TUM School of Computation, Information and Technology
Advisor:
Günnemann, Stephan (Prof. Dr.)
Referee:
Günnemann, Stephan (Prof. Dr.); Bronstein, Michael (Prof., Ph.D.)
Language:
en
Subject group:
DAT Datenverarbeitung, Informatik
Keywords:
Machine learning, graph, molecule, graph neural network
Translated keywords:
Maschinelles Lernen, Graph, Molekül, Graph-neuronale Netze
TUM classification:
DAT 600; DAT 703; DAT 708
Abstract:
Graph neural networks (GNNs) have recently enabled great advances in leveraging graphs for accurate predictions. However, regular GNNs ignore the fact that graphs are often embedded in an underlying geometrical space. This thesis aims at alleviating this limitation by proposing and analyzing methods that go beyond structure and incorporate geometric information such as distances and directions. This enables models that are more accurate and robust, generalize better, and scale to larger graphs.
Translated abstract:
Graph-neuronale Netze (GNNs) haben große Fortschritte darin ermöglicht, Graphen für maschinelle Vorhersagen zu nutzen. Jedoch ignorieren reguläre GNNs, dass der sichtbare Graph oft in einen geometrischen Raum eingebettet ist. Diese Dissertation erweitert deshalb GNNs durch Methoden, die strukturelle Informationen mit geometrischen Informationen wie Distanzen und Richtungen verbinden. Dies ermöglicht Modelle, die präziser und robuster sind, besser generalisieren und zu großen Graphen skalieren.
WWW:
https://mediatum.ub.tum.de/?id=1660288
Date of submission:
12.07.2022
Oral examination:
03.03.2023
File size:
6644541 bytes
Pages:
210
Urn (citeable URL):
https://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:91-diss-20230303-1660288-1-9
Last change:
27.04.2023
 BibTeX