Practice English vocabulary for graph machine learning: GNNs, GCNs, node embeddings, link prediction, knowledge graph completion, and embedding models like TransE.
0 / 5 completed
1 / 5
What does a Graph Neural Network (GNN) do that a standard neural network cannot?
A GNN iteratively aggregates messages from neighbouring nodes to update each node's representation. This makes it possible to learn from both node features and graph topology simultaneously — something flat feature vectors cannot express.
2 / 5
How would you describe 'node embedding' to a data scientist unfamiliar with graphs?
Node embeddings (e.g., from Node2Vec or GNN encoders) represent each node as a real-valued vector. The key property is that geometric proximity in the embedding space reflects graph structure or semantic similarity, enabling downstream ML tasks.
3 / 5
What is 'link prediction' as a graph ML task?
Link prediction asks: given the current graph, which pairs of nodes are likely to be connected by an edge that is missing? It is fundamental to recommender systems (user–product links), social networks (friend suggestions), and KG completion (missing facts).
4 / 5
What problem does 'knowledge graph completion' address?
Real-world knowledge graphs like Wikidata are incomplete — many true facts are missing. KG completion models (TransE, RotatE, ComplEx) learn entity and relation embeddings and use them to score and predict missing (subject, predicate, object) triples.
5 / 5
What is the key idea behind the TransE knowledge graph embedding model?
TransE (Translating Embeddings) is an intuitive KG embedding model: h + r ≈ t for valid triples. It is efficient and interpretable, though it struggles with symmetric and one-to-many relations. RotatE and ComplEx extend it to handle these cases.