Overview of Graph Embedding Techniques: DeepWalk, LINE, node2vec, and EGES

1. Introduction to Graph Embedding

Graph Embedding extends Word2Vec‑style sequence embedding to graph‑structured data, generating low‑dimensional dense vectors for nodes that capture structural and local similarity.

2. DeepWalk

DeepWalk performs random walks on an item graph to produce node sequences, which are fed into Word2Vec to learn embeddings. The original paper “Billion‑scale Commodity Embedding for E‑commerce Recommendation in Alibaba” describes the workflow, including graph construction from user behavior and transition probability definitions for directed weighted graphs.

3. LINE – Improving DeepWalk

LINE replaces DFS‑based walks with BFS‑style neighborhood construction and defines first‑order and second‑order proximity. First‑order proximity models direct edge weights, while second‑order proximity models shared neighborhoods, using a weighted KL‑divergence objective and negative sampling.

4. node2vec – Balancing Homophily and Structural Equivalence

node2vec introduces biased second‑order random walks controlled by return parameter p and in‑out parameter q , allowing interpolation between breadth‑first (BFS) and depth‑first (DFS) strategies to capture both homophily and structural equivalence in the learned embeddings.

5. EGES – Enhanced Graph Embedding with Side Information

EGES extends the Base Graph Embedding (BGE) pipeline by incorporating side information (category, brand, price) with learnable weights, using a weighted skip‑gram model to alleviate cold‑start and sparsity issues in large‑scale e‑commerce recommendation.

The recommendation pipeline typically consists of matching and ranking stages, where graph embeddings are used to compute item similarity in the matching phase.

6. Summary

Graph Embedding remains a hot research and industry topic; besides DeepWalk, LINE, node2vec, and EGES, methods such as SDNE and struct2vec also merit further study.