1. Foundational Papers (Must Read)
| Year | Paper | Why it matters |
|---|---|---|
| 2014 | Bruna - Spectral Networks | First spectral GNN - started the field |
| 2016 | Defferrard - ChebNet | Localized spectral filters |
| 2017 | Kipf & Welling - GCN | The model everyone starts with |
| 2017 | Gilmer - MPNN | Unified message passing framework |
| 2017 | Hamilton - GraphSAGE | Inductive learning at scale |
| 2018 | Veličković - GAT | Attention on graphs |
| 2019 | Xu - GIN / WL power | Expressiveness upper bound |
2. Benchmarks & Ecosystem
- Hu et al. - Open Graph Benchmark (OGB, 2020) - standardized datasets and evaluators
- Bronstein et al. - Geometric Deep Learning Blueprint (2021) - theoretical unification
- PyTorch Geometric - main research library
- DGL - scalable graph deep learning
3. Expressiveness & Limits
4. Graph Transformers Era
5. Self-Supervised & Foundation Models
- Hou et al. - GraphMAE (2022) - masked autoencoder for graphs
- Liu et al. - GraphGPT (2023)
- Liu et al. - OneForAll (2023) - one model many graph tasks
- GFM survey (2024) - graph foundation models overview
6. Molecules & Equivariant GNNs
7. LLM + Graph (2024–2026)
- Edge et al. - GraphRAG (Microsoft, 2024)
- Chen et al. - Text-Attributed Graphs survey
- Think-on-Graph (2024) - LLM reasoning with KG
- G-Retriever (2024) - RAG with textual graphs
- GraphLLM (2024)
Full index: Papers Hub
8. Theses & Books
- Graph Representation Learning - William L. Hamilton (book, free online) - best textbook, read cover to cover
- Hamilton PhD thesis (Stanford) - GraphSAGE and large-scale learning
- Xu Keyulu PhD thesis (MIT) - expressiveness of GNNs, GIN
- Geometric Deep Learning - Bronstein, Bruna, Cohen, Veličković (geometricdeeplearning.com)
- Graph Neural Networks: Foundations, Frontiers, and Applications - Zhou et al. (Springer)
9. Open Problems
- Scaling laws for GNNs - do they follow LLM-style power laws?
- True graph foundation models - one pretrained model for any graph domain?
- Over-squashing cure - practical fix without $O(n^2)$ attention
- Heterophily at scale - production models on web-scale heterophilous graphs
- Dynamic graphs + LLMs - streaming KG updates with frozen LLM
- Theory of depth - principled layer count selection
- Fairness and bias on graph recommendations
- Explainability - which subgraph caused the prediction?