GenAIWiki
Machine learning

Graph Machine Learning

Graph machine learning applies statistical and neural methods to data represented as nodes, edges, and attributes so models can learn from relationships as well as individual records.

Expanded definition

Graph machine learning is the set of methods used when relationships are part of the data rather than incidental metadata. A graph represents entities as nodes, relationships as edges, and optional properties as attributes. Practitioners use graph methods for node classification, link prediction, graph classification, recommendations, fraud detection, knowledge graphs, and network analysis. Graph neural networks (GNNs) commonly update a node representation by aggregating information from neighboring nodes. Other graph workflows use embeddings, community detection, centrality measures, or path analysis without requiring a neural network. The useful abstraction is relational structure: the same table of entities can support different predictions once its connections are modeled explicitly. GraphRAG is a retrieval pattern adjacent to graph machine learning. Microsoft GraphRAG extracts entities, relationships, and claims from source documents, detects communities, generates community reports, and uses those graph-derived structures during retrieval. This can help with questions that require connecting facts across a corpus, but it adds indexing work and operational cost. Standard chunk-based RAG remains the simpler default for direct evidence retrieval.

Common graph learning tasks

The prediction target determines the graph representation, training data, and evaluation metric. Many production systems combine graph features with conventional tabular or text features.

FamilyHow it worksExamplesCommon outputs
Node classificationPredict a label for a node using its attributes and neighborhood.Account risk scoring, document topic labelingNode label or probability
Link predictionEstimate whether an edge exists or should exist between two nodes.Recommendations, fraud-ring discoveryEdge score or ranked candidate links
Graph classificationPredict a label or property for an entire graph.Molecule property prediction, program analysisGraph label or regression value
Community detectionGroup densely connected nodes to expose clusters and higher-level structure.Network segmentation, GraphRAG community reportsCommunity assignments and summaries

GraphRAG vs standard RAG

DimensionStandard RAGGraphRAG
Indexing pathChunk documents, create embeddings, and index chunks with metadata.Extract entities, relationships, and claims; detect communities; generate community reports; embed graph-derived text artifacts.
Retrieval strengthDirect passage retrieval for questions answered by a small set of relevant chunks.Connected evidence and corpus-level themes that benefit from entity neighborhoods or community summaries.
Query modesVector, keyword, or hybrid retrieval followed by generation.Local search combines graph data with source text for entity-focused questions; global search uses community reports for dataset-wide reasoning.
Operational profileSimpler ingestion, lower indexing overhead, and easier debugging.More indexing stages, prompt-driven extraction, summary generation, and graph artifacts to inspect.
Best fitSupport search, documentation Q&A, and grounded answers where retrieved passages are sufficient.Research and analysis questions that require relationships, multi-hop evidence, or high-level themes across a corpus.

Start with standard RAG and a measured evaluation set. Add GraphRAG when errors are caused by fragmented evidence, relationship traversal, or corpus-level questions rather than weak chunking or ranking. Treat the graph index as an additional retrieval system with its own quality, latency, and cost budget.

Related terms

Explore adjacent ideas in the knowledge graph.

Related

Comparisons, tools, and models that connect to this idea.