Graph Neural Networks for Link Prediction with Subgraph Sketching

Explore a comprehensive talk on Graph Neural Networks (GNNs) for Link Prediction, focusing on innovative approaches to overcome limitations in expressive power and efficiency. Learn about the challenges faced when using GNNs for link prediction tasks, including their inability to count triangles and distinguish automorphic nodes. Discover how subgraph-based methods have achieved state-of-the-art performance but suffer from poor efficiency. Delve into the analysis of subgraph GNN components and the novel full-graph GNN called ELPH (Efficient Link Prediction with Hashing), which uses subgraph sketches to approximate key components without explicit subgraph construction. Understand the development of BUDDY, a highly scalable model that uses feature precomputation to overcome memory limitations. Gain insights into the performance and efficiency of these new approaches compared to existing models on standard Link Prediction benchmarks.

- Intro
- Production Model at Twitter
- First Attempt with GNNs for Recommendations
- Automorphic Nodes
- Subgraph GNNs
- Analyzing Subgraph GNNs: Model Components
- Subgraph Sketching: Minhash
- GNNs with Subgraph Sketching
- BUDDY: Scaling ELPH
- Results
- Q+A