Graph & Network ML #3

Node2Vec / DeepWalk

  • DeepWalk
  • Node2Vec
  • Random-Walk Embeddings

Best for: Graph embeddings Aliases: DeepWalk, Node2Vec, Random-Walk Embeddings

How it works

$$\max_\theta\sum_u\sum_{v\in N(u)}\log p(v\mid u;\theta)$$

Generates many random walks over the graph — in Node2Vec biased by parameters $p,q$ that interpolate between BFS and DFS exploration — and feeds the walk sequences to a skip-gram model. The objective maximises $\sum_u\sum_{v\in N(u)}\log p(v\mid u;\theta)$ over learned node embeddings $z_u\in\mathbb R^d$. The resulting low-dimensional vectors preserve structural proximity and serve as features for downstream tasks such as node classification and link prediction.

When to use

Generating node embeddings for downstream ML when you need simple, scalable feature learning.

Watch out

Random-walk parameters shape results; ignores edge/feature updates; superseded by GNNs for many tasks.

Common fields

Recommendations · link prediction