DataTerminology.comNEW
Also known as: Edge prediction
Predicting missing, future, or likely relationships between nodes in a graph using structural and feature information.
DESCRIPTION
Link prediction estimates whether a relationship exists or will form between two nodes. Classical heuristics include common neighbors and Adamic–Adar; modern methods use embeddings and GNNs. It powers recommendations, fraud detection, and knowledge graph completion. Evaluation uses temporal splits, negative sampling, and metrics like AUC or Hits@K. Robust pipelines manage leakage by preserving temporal order and separating overlapping neighborhoods.
ALTERNATIVE DEFINITIONS
Wikipedia
Link prediction is the problem of predicting the existence of a link between two nodes in a network.
Concise problem statement focusing on link existence.
Stanford SNAP (CS224W)
Link prediction aims to infer missing edges or forecast future interactions using network topology and node features.
Explicates both missing and future links and feature use.
arXiv
The link prediction task asks whether two nodes should be connected, based on structural patterns and learned representations.
Highlights role of learned representations in modern approaches.
RELATED TERMS
Graph Neural Network Graph Embedding Similarity Measure Common Neighbors Preferential Attachment Knowledge Graph Completion Node2Vec Graph Convolutional Network Matrix Factorization Recommendation System
COMPETING TERMS
RELATIONS
- Broader: Prediction
KNOWLEDGE AREAS
AI Data Science Semantic & Knowledge Modeling
TARGET AUDIENCE
Data Scientist AI Engineer MLOps Engineer Data Modeler
EXTERNAL SOURCES
- Wikipedia - Link prediction
- Stanford SNAP - Course notes on link prediction
- PyTorch Geometric - Link prediction tutorial
- Deep Graph Library - Mini-batch link prediction
TOP QUESTIONS
What are baseline heuristics for Link Prediction?
Common neighbors, Jaccard, Adamic–Adar, and preferential attachment provide quick, interpretable baselines.
Category: Methods
How is Link Prediction evaluated?
Use temporal splits, negative sampling, and metrics like AUC, MAP, or Hits@K.
Category: Metrics
Why is temporal validation important for Link Prediction?
It prevents leakage by ensuring future links aren’t predicted with future information.
Category: Best Practices
How do GNNs improve Link Prediction?
GNNs aggregate neighborhood information and learn representations capturing higher-order structure.
Category: AI
What is negative sampling in Link Prediction?
Sampling non-existent edges as negatives for supervised training, balanced to reflect graph sparsity.
Category: Concepts
What pitfalls occur with highly skewed graphs in Link Prediction?
Class imbalance and popularity bias can inflate scores; calibration and stratified sampling help.
Category: Risks
Where is Link Prediction used in industry?
Friend suggestions, product recommendations, fraud ring detection, and knowledge graph completion.
Category: Applications
How does edge type affect Link Prediction?
Heterogeneous graphs require type-specific features or meta-paths to capture semantics.
Category: Modeling
What is the role of node features in Link Prediction?
Features complement topology, improving prediction when structure alone is insufficient.
Category: Concepts