Toward Universal Graph Representations: Foundations and Frontiers of Graph Foundation Models
summary
Proposes concrete recipe: create holographic node representations via an expansion map that breaks symmetries (Lanczos-style or GNN-initialized perturbations) then a reduction map that restores only the task’s equivariance, which provably enables k‑node joint predictors that standard GNNs cannot and empirically closes large performance gaps in transfer between node, edge and higher‑order tasks. Introduces a feature hypergraph construction and hypergraph GNN that encodes distribution‑free permutation invariances for zero‑shot transfer across disparate feature domains and shows improved accuracy on e‑commerce benchmarks while noting computational and expressivity tradeoffs.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:29Graph foundation model framework
- 5:46Holographic node representations
- 13:59Why GNNs fail for joint predictions
- 24:58Holographic node representations explained
- 32:46Symmetries in GNNs
- 42:36Feature domain transferability
- 49:18Zero-shot learning across domains
- 59:43Future of graph foundation models
related talks
01h 03mTowards Foundation Models for Graphs
16mPreview: Graph NNs
39mUsing Graph Neural Networks to Embrace the Dependency Within Your Data
18mGraphon Neural Networks and the Transferability of Graph Neural Networks
34mImproving Data Understanding with Graphical Hypergeometric Networks
59mTransforming DataFrames: Intro to Graphs & GNNs with PyTorch Geometric
31mGraph-Based Machine Learning with Python
36mAmazing things that happen with Human-AI synergy