Using Graph Neural Networks to Embrace the Dependency Within Your Data
summary
Claim: graph neural networks beat naive tabular models when relationships matter by encoding structure into adjacency and Laplacian matrices and learning via message passing aggregations (sum/mean/max). Concrete example: bike‑share demand predicted as a node task using node features plus edges that encode trips, distance or transit. Implementation notes: sparse ops avoid huge zero matrices, too many GNN layers “blur” signals, and production stacks commonly use Neo4j, Spark GraphX, PyTorch Geometric (±PyG‑Temporal). Papers from 2017–2018 report significant improvements over non‑graph baselines.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 2:39Problem solving
- 3:29Example use-case
- 8:41Framing the problem...using a graph
- 9:30Encoding information in the edges
- 12:53Types of Problems
- 16:39Node Prediction
- 16:58Adjacency Matrix
- 18:01Diagonal Degree Matrix
- 18:30Laplacian
- 18:50Edge Features
- 21:35Message Passing
- 22:36Mathematically...
- 24:21End-To-End Process
- 26:32Performance
- 27:18Graph Database
- 29:12Spark Graphx
- 29:46PyTorch Geometric (Temporal)
related talks
16mPreview: Graph NNs
31mGraph-Based Machine Learning with Python
01h 01mToward Universal Graph Representations: Foundations and Frontiers of Graph Foundation Models
29mGraph Neural Networks + LLM on
59mTransforming DataFrames: Intro to Graphs & GNNs with PyTorch Geometric
19mNeuGraph: Parallel Deep Neural Network Computation on Large Graphs
01h 16mBoot camp on graph foundations models
36mAmazing things that happen with Human-AI synergy