LINE (Large-scale Information Network Embedding) is a graph embedding method designed explicitly for massive networks (millions of nodes) that learns node representations by optimizing two complementary proximity objectives — first-order proximity (connected nodes should be close) and second-order proximity (nodes sharing common neighbors should be close) — using efficient edge sampling to achieve linear-time training on billion-edge graphs.

What Is LINE?

• Definition: LINE (Tang et al., 2015) learns node embeddings by separately optimizing two objectives: (1) First-order proximity preserves direct connections — the embedding similarity between two connected nodes should match their edge weight: $p_1(v_i, v_j) = sigma(u_i^T cdot u_j)$ where $sigma$ is the sigmoid function. (2) Second-order proximity preserves neighborhood overlap — nodes sharing many common neighbors should have similar embeddings, modeled by predicting the neighbors of each node from its embedding using a softmax: $p_2(v_j mid v_i) = frac{exp(u_j'^T cdot u_i)}{sum_k exp(u_k'^T cdot u_i)}$.
• Separate then Concatenate: LINE trains two sets of embeddings — one for first-order and one for second-order proximity — then concatenates them to form the final embedding vector. This separation avoids the difficulty of jointly optimizing two different structural signals and allows independent tuning of each proximity's embedding dimension.
• Edge Sampling: To avoid the expensive softmax normalization over all nodes, LINE uses negative sampling (sampling random non-edges) and alias table sampling for efficient edge selection — enabling stochastic gradient descent with $O(1)$ cost per update rather than $O(N)$ for full softmax.

Why LINE Matters

• Scale: LINE was the first embedding method explicitly designed for billion-scale graphs — its edge sampling strategy enables training on graphs with billions of edges in hours on a single machine. DeepWalk's random walk generation and Node2Vec's biased walks both have higher per-edge overhead than LINE's direct edge sampling.
• Explicit Proximity Decomposition: LINE's separation of first-order (direct connections) and second-order (shared neighborhoods) proximity provides a clean framework for understanding what graph embeddings capture. First-order proximity encodes the local edge structure; second-order proximity encodes the broader neighborhood pattern. Different downstream tasks benefit from different proximity types.
• Directed and Weighted Graphs: LINE naturally handles directed and weighted graphs — the asymmetric second-order objective models directed edges by using separate source and context embeddings, and edge weights directly modulate the training gradient. DeepWalk and Node2Vec require additional modifications for directed or weighted graphs.
• Industrial Adoption: LINE's simplicity, scalability, and explicit objectives made it one of the most widely deployed graph embedding methods in industry — used for recommendation systems (embedding users and items from interaction graphs), knowledge graph completion, and large-scale social network analysis.

LINE vs. Other Embedding Methods

LINE is explicit proximity mapping — directly forcing connected nodes and structurally similar nodes to align in vector space through two clean, complementary objectives, achieving industrial-scale graph embedding through the simplicity of edge-level optimization rather than walk-level sequence modeling.