Format des notes
Numérique sur 20
Littérale/grade européen
Pour les étudiants du diplôme Diplôme d'Ingénieur de l'Ecole Nationale Supérieure de Techniques Avancées (Institut polytechnique de Paris)
Le rattrapage est autorisé (Note de rattrapage conservée)- le rattrapage est obligatoire si :
- Note initiale < 7
- le rattrapage peut être demandé par l'étudiant si :
- 7 ≤ note initiale < 10
L'UE est acquise si Note finale >= 10- Crédits ECTS acquis : 3 ECTS
Programme détaillé
Schedule:
- Lecture 1: Introduction to Machine Learning on Graphs.
- Motivation: Why graphs matter in ML (examples from social networks, biology, internet, chemistry).
- Graph basics: nodes, edges, directed/undirected graphs, etc.
- Applications of graph learning (social networks, molecules, recommender systems).
- Classical ML on graphs: centrality measures, node embeddings (DeepWalk, Node2vec).
- Introduction to graph representation learning.
- Lecture 2: Introduction to Graph Neural Networks.
- Neural networks on structured data.
- Convolutions in time (1D), space (2D), and graphs.
- Graph signals and convolutional filters.
- Graph Convolutional Networks (GCNs).
- Message Passing Neural Networks (MPNN) framework.
- Graph Attention Networks (GAT).
- Limitations of GNNs (over-smoothing, heterophily).
- Introduction to PyTorch Geometric (PyG).
- Lab Session 1: Graph Neural Networks.
- Lecture 3: Scaling up Graph Neural Networks.
- Challenges with large-scale graphs: memory, mini-batch training, neighborhood explosion.
- Applications with massive graphs: recommender systems, social networks, citation networks.
- Why SGD is not straightforward for GNNs.
- Scalable GNN approaches:
- GraphSAGE (neighbor sampling).
- Cluster-GCN (graph partitioning).
- Simplifying GNN architectures (SGC).
- Trade-offs: scalability vs. accuracy.
- Lecture 4: Spatiotemporal Analysis with Graph Neural Networks.
- Introduction to spatiotemporal data and time-varying graph signals.
- Temporal operators and difference signals.
- Spectral GNNs.
- Spatiotemporal graph neural networks.
- Applications: traffic prediction, weather forecasting.
- Lab Session 2: Spatiotemporal Analysis with Graph Neural Networks.
- Lecture 5: Recommender systems.
- Bipartite graphs: users and items.
- Task: predicting new edges (future interactions).
- Embedding-based models for recommendation.
- Evaluation metrics: Recall@K, NDCG.
- GNNs for recommendation: user–item embedding with message passing.
- Lecture 5: Graph Generation.
- Applications: drug discovery, material design, social network modeling, anomaly detection.
- Properties of real-world graphs (degree distribution, clustering coefficient, etc.).
- Traditional graph generative models:
- Erdős–Rényi random graphs.
- Deep generative models for graphs:
- GraphRNN.
- Efficient graph generation methods.
- Lab Session 6: Graph Generation.
- Exam.
