Boot camp on graph foundations models
summary
Claims that pretrained graph models give real transfer: a relation-graph encoder + labeling tricks (NBFNet, motif hierarchy) yields zero-shot link prediction that beats per-dataset training by about 10% on average and runs from ~100 to 1M nodes; a tiny attention head over analytic linear predictors (the GraphAny linear GNN ensemble) classifies graphs with 3–200 classes and often outperforms end-to-end GNNs zero-shot; soft-token graph encoders and a AgentsNet multi-agent benchmark show LLM agents can reach consensus but fail on harder coloring/cover tasks as network size grows.
AI-written summary · about
chapters (11)
- 0:00Introduction and overview
- 2:16Graph mining at Google
- 3:38Graph foundation models
- 10:30Link prediction success
- 22:45Node classification methods
- 27:56Challenges in graph modeling
- 30:40Foundation models for databases
- 35:09LLMs for graph foundation
- 43:58Encoding graphs for LLMs
- 51:43AgentsNet and collaborative AI
- 1:02:55Q&A and concluding remarks
related talks
16mPreview: Graph NNs
01h 03mTowards Foundation Models for Graphs
39mUsing Graph Neural Networks to Embrace the Dependency Within Your Data
31mGraph-Based Machine Learning with Python
23mUnveiling the Journey of NLP
18mEvaluating the effectiveness of Deep Learning Models for Foundational Program Analysis(…)
29mGraph Neural Networks + LLM on
12mMake ML outcomes roar with graph neural networks on TigerGraph and AWS (PRT097)