Graph Neural Networks + LLM on
summary
Demonstration of GNN+LLM pipelines shows PyTorch Geometric 2.7 integrated with Neo4j for GraphRAG workflows, including a text-to-KG class that extracts triples from documents, a PCST 2-approx pruning step, and end-to-end training where GNN embeddings project into LLM token space. Benchmarks on the STARKPrime biomedical set (129k nodes, 8M edges) double prior leaderboards from hits@1 15%→32%, run on modest GPUs (~4 GB per model), and ship with examples for PyG 2.7 and Neo4j.
AI-written summary · about
chapters (7)
- 0:00Introduction and agenda
- 1:39Limitations of autoregressive
- 2:20GNN and GraphRAG fundamentals
- 7:14Building the GraphRAG pipeline
- 16:16Knowledge graph creation
- 17:13Performance and future outlook
- 22:21Q&A and closing remarks
related talks
23mGNN-RAG: Bridging Graph Reasoning and Language Understanding for Knowledge Retrieval
41mEnhancing LLMs with Graph Technology
31mGraph-Based Machine Learning with Python
46mKnowledge Graphs + LLMs: What's the latest in research?
28mGrounded Intelligence Boosting LLMs with Real-Time Graph Context
53mPractical GraphRAG - Making LLMs smarter with Knowledge Graphs
16mPreview: Graph NNs
36mGraphRAG Explained 🔥 Building Smarter LLMs