The mini-course covers the basics of Graph Representation Learning / Graph Neural Networks.
The following topics are included:
- Motivation for graph representation learning. Several examples of benchmarks.
- Brief review of graph and matrix theory, notion of embeddings.
- Necessary NLP models reminder: from word2vec to Transformers.
- Random Walk Graph Embeddings: node2vec, struct2vec.
- Message Passing Embeddings: GraphSAGE, Graph Attention Network.
- Unsupervised Graph Embeddings, notion of Triplet Loss.
Repository is organized as follows:
- Graph_Slides.pdf - materials for the lectures
- Seminars:
- PyTorch_Geometric_Dataset.ipynb - introduces CORA dataset, covers the creation of PyTorch Geometric Dataset based on this graph. This dataset will be used in the following seminars for model training.
- GraphModel_Supervised.ipynb - acquaints with NeighborSampler and the concept of mini-batch learning on large graphs. Then the training process of supervised Graph Attention Network is covered.
- GraphModel_Unsupervised.ipynb - covers training of unsupervised GraphSAGE model. Visualization of resulting node embeddings using UMAP is included.
We recommend to download ipynb files for correct/proper display of HTML elements.
