This course offers an in-depth introduction to graph learning, combining theoretical foundations with state-of-the-art applications. The first module covers fundamental concepts in graph theory, with a focus on formal definitions, topological graphs, and essential structural properties. The second module delves into graph neural networks (GNNs), exploring both spectral methods, giving few insights of graph signal processing, and spatial approaches. Classic graph learning tasks such as node classification, link prediction, on standard dataset of citations and publication networks will be investigated with additional colab sessions. The final module examines advanced applications of graph learning beyond traditional tasks. Topics include graph-based techniques in computer vision and machine learning with discussion of recent publications and case studies. The final assessment will consist of a short report outlining how graph learning techniques can be applied within the student’s own PhD research.

Basic knowledge of probability and statistics, linear algebra and calculus Basic knowledge of machine learning and deep neural networks Experience with python, pytorch

1. Introduction to graphs (1h) 2. Node embedding (2h) 3. Graph neural networks (3h): - Connection to CNNs, Fourier transform - Spectral graph theory, spectral GNNs - Message passing, spatial GNNs - Graph pooling 4. Classic GNNs problems (2h): - GNNs for node classification + colab session - GNNs for graph classification + colab session - GNNs for link prediction + colab session 5. Graph Generation (2h) 6. Advanced applications (10h) - Geometric deep learning and dynamical systems for graph processing - Deep learning for dynamic graphs and spatio-temporal data - Graph learning for video understanding

In presenza

Presentazione report scritto

P.D.2-2 - Ottobre