Back
2025INF001
Doctoral thesis

Geometric deep learning : from grid to graph structured data

  • 2025

PhD: Università della Svizzera italiana

Geometric deep learning
Graph convolutional neural networks
Graph-structured data
Neutrino detection
Misinformation detection
User identification
English The success of Deep Learning architectures (e.g. Convolutional Neural Networks, Recurrent Neural Networks, Transformers, ...) and the increasing availability of graph/manifold structured data (e.g. social networks, sensor networks, molecules, 3D shapes, ...) motivated, in recent years, the development of a new class of Geometric Deep Learning (GDL) approaches aimed at extending traditional DL solutions to non-Euclidean domains. In this thesis, we explore the realm of Graph Convolutional Neural Networks (GCNNs), a popular class of GDL architectures that rely on generalizations of the convolution operation to process the provided input data. Our contributions are organized in two main different parts. In the first part of this manuscript, we focus on methodologies. We introduce in particular novel generalizations of convolution that are defined either in space or in the spectral domain. We present an attention mechanism able to generalize convolution through a generalization of the notion of pixel (MoNet), spectral filters able to process signals defined over multiple graphs (MGCNN), spectral filters with spectral zoom properties (CayleyNet), and a scalable inception-based architecture able to efficiently process graphs with millions of nodes and billions of edges (SIGN). In the second part of our work, we direct our attention towards applications of GCNNs. First, we show how GCNNs allow to detect high-energy neutrinos by processing signals retrieved by the IceCube detector. Second, we introduce VeritasZero, a GCNN-based approach able to detect fake news based on the spreading patterns this forms on social media. Third, we present BPIIG, a GCNN-based profiling attack that exploits the stability of people’s interaction behavior to identify individuals in anonymous datasets.
Collections
Language
  • English
Classification
Computer science and technology
License
License undefined
Open access status
green
Identifiers
Persistent URL
https://n2t.net/ark:/12658/srd1330972
Statistics
Document views: 49 File downloads:
  • 2025INF001.pdf: 42