ML Katas

Implement a Graph Autoencoder (GAE) for graph representation learning. The encoder will use a GNN to produce node embeddings, and the decoder will reconstruct the graph's adjacency matrix from these embeddings. The goal is to learn a low-dimensional representation of the nodes that preserves the graph's structure. You can use a simple GNN layer from a library like torch-geometric or implement your own from scratch.

Verification: After training, take the learned node embeddings and use them for a downstream task like link prediction (predicting missing edges). A high AUC score for link prediction indicates that the autoencoder has successfully learned a good representation of the graph's structure.