Graph Neural Networks (GNNs) have emerged as a major area of machine learning research, powering advances in domains from drug discovery to recommendation systems. But diving into GNN research can be overwhelming given the sheer volume and diversity of papers published over the past decades. This is where the gnnpapers repository maintained by the THUNLP group at Tsinghua University offers a rare, practical resource: a meticulously curated bibliography of over 800 essential GNN papers.
A comprehensive curated bibliography for graph neural networks
Rather than providing code or models, this repo serves as a structured reading list that traces the evolution and breadth of graph neural network research. It assembles papers from foundational models, surveys, efficiency and explainability techniques, all the way to 18 application domains including drug discovery, knowledge graphs, traffic forecasting, and program representation.
The taxonomy is organized thoughtfully. It starts from early structural classification work by Sperduti in 1997, through seminal graph convolutional network (GCN) models by Kipf and Welling, and extends to modern self-supervised and dynamic graph approaches. This makes the repo not only a reference but also a historical map of how the field has developed.
Despite having no executable code or implementations, the repo has amassed nearly 17,000 stars on GitHub. This underlines the hunger in the community for a single, authoritative source to navigate GNN literature, especially for newcomers or those branching into this specialized topic.
What makes this bibliography stand out and the tradeoffs involved
The key strength of this repo is its exhaustive scope combined with community recognition and maintenance by a reputable academic group. It is not just a flat list but a carefully categorized taxonomy that balances breadth and depth, which is crucial when faced with a fast-evolving research area.
The curated metadata includes categorization by types of GNN models — spectral versus spatial convolutions, message passing frameworks, graph attention mechanisms — and by application domains. This makes it easier to identify relevant papers based on research interests or practical needs.
However, the tradeoff is obvious: the repo contains zero code. For practitioners looking to quickly prototype or benchmark GNN models, this resource alone won’t suffice. Instead, it complements code-centric repos by offering a clear path to the original research publications that underpin the models and techniques implemented elsewhere.
The quality of curation also means that the repo requires active maintenance to keep pace with new publications. Its community-driven nature helps here, but there is always a lag between cutting-edge research and inclusion in the list.
Explore the project: navigating the GNN papers repository
Since the repo focuses on bibliography and does not provide software or installation instructions, the best way to use it is to explore its structure and documentation.
The main entry point is the README, which outlines the taxonomy and provides links to categorized lists of papers. These are usually organized in markdown files or directories named by topic or application domain.
For example, you will find separate sections or files for foundational survey papers, spectral GNN models, spatial models, attention-based methods, and various applications like drug discovery or traffic forecasting. Each paper entry typically includes the title, authors, publication venue, and a link to the original paper.
The repo also sometimes links to community resources or highlights landmark papers that are essential reading.
If your interest is in a specific GNN subfield or application, start by scanning the relevant section. For newcomers, beginning with the survey and foundational models sections is recommended to build a solid theoretical base.
Verdict: who should bookmark gnnpapers and its limitations
This repo is a must-have bookmark for anyone seriously involved in graph neural network research or applications. It saves the painful time otherwise spent hunting down relevant papers scattered across conferences and journals.
Academics, PhD students, and research engineers will find it invaluable for literature reviews and staying updated on GNN trends. It also benefits machine learning practitioners who want to deepen their understanding beyond code libraries.
The main limitation is the lack of runnable examples or implementations. Users will need to complement this resource with code repos and frameworks to build or experiment with GNN models.
In summary, gnnpapers exemplifies how well-organized knowledge curation can compound value in a technical field. It stands as a testament that sometimes the most impactful open source projects aren’t codebases but living maps of knowledge that guide practitioners through complex research landscapes.
Related Articles
- Automating knowledge graph extraction from text with LangChain and GPT-4o — This repo uses LangChain’s experimental graph transformers with GPT-4o to extract and visualize knowledge graphs from un
- AI Knowledge Graph Generator: Building structured graphs from unstructured text with LLMs — A Python tool that converts unstructured text into interactive knowledge graphs using a three-phase LLM pipeline with SP
- graph-memory: a knowledge graph context engine plugin for AI agents using dual-path recall — graph-memory compresses AI agent context by 75% using a knowledge graph and dual-path recall with Personalized PageRank.
- annotated_deep_learning_paper_implementations: annotated PyTorch implementations of key deep learning papers — This repo provides annotated PyTorch implementations of major deep learning papers with side-by-side explanations, aidin
- OpenResearcher: An open-source 30B LLM for long-horizon deep research — OpenResearcher is a fully open 30B agentic LLM designed for deep research tasks, featuring a 96K-turn dataset and a self
→ GitHub Repo: thunlp/GNNPapers ⭐ 16,780