Module node2vec 

Node2Vec: Scalable Feature Learning for Networks

Reference: Grover & Leskovec (KDD 2016) — https://arxiv.org/abs/1607.00653

Algorithm Overview

Node2Vec generates node embeddings via biased second-order random walks. Two hyperparameters control the walk bias:

• p (return parameter) – probability of returning to the previous node. High p → DFS-like walk (explores further from source).
• q (in-out parameter) – probability of moving away from the previous node. Low q → BFS-like walk (stays near source).

The algorithm:

1. Pre-compute per-edge alias tables for O(1) biased sampling.
2. Simulate num_walks second-order random walks of length walk_length from every node.
3. Train a simplified skip-gram model on the walk corpus to produce embedding_dim-dimensional embeddings.

Implementation Notes

• No external ML or linear-algebra crates are required.
• Random number generation uses scirs2_core::random per project policy.
• All unsafe code is avoided; heavy loops use Rust iterators.

Structs

Node2VecConfig

Full Node2Vec configuration (walks + skip-gram training).

Node2VecEmbedder

Node2Vec graph embedding generator.

Node2VecEmbeddings

Computed Node2Vec embeddings.

Node2VecWalkConfig

Configuration for the random-walk phase of Node2Vec.