graph-based-deep-learning-literature: A Curated Index of Graph Deep Learning Conference Papers
links to conference publications in graph-based deep learning
At a glance
- What is it?
- The repository is a link collection, not a library. It maps NeurIPS, ICML, ICLR and KDD graph papers into year and topic folders, and its usefulness depends entirely on whether you want a reading list rather than runnable code.
- Who is it for?
- Adopt it if you need a conference-scoped reading list for graph representation learning and you are willing to follow links out to arXiv or OpenReview for the actual papers. Do not adopt it if you want runnable implementations, benchmark numbers or a maintained library: the README points to a separate software and libraries page for that.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 3 days ago.
- What is it written in?
- Mainly Jupyter Notebook, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on October 1, 2026, and from our analysis. They are not legal advice.
Editorial analysis
What the repository actually is, and who it is for
The README states the repository "primarily contains links to conference publications in graph-based deep learning." That sentence is the whole product. There is no Python package, no importable module and no CLI. A researcher who wants to know what appeared at ICLR 2024 on graph representation learning opens a folder and reads a list. A researcher who wants to train a model on a citation network is in the wrong place.
The audience is narrow but real. Doctoral students doing a literature review, reviewers checking whether a claimed contribution is novel against a conference's recent output, and engineers who need to know which ideas have academic traction before committing to an architecture. The topics list on the repository confirms the scope: graph convolutional networks, graph neural networks, graph representation learning. Everything else is out of scope by design.
The organising principle is conference first, year second, topic third. That matters more than it sounds. A topic-first index tells you what exists. A conference-first index tells you what the field's gatekeepers accepted, which is a different and often more useful signal when you are deciding what to read.
How the folder structure encodes the data model
The top level holds three entries: LICENSE.md, README.md and conference-publications/. Everything else lives under that third directory. Inside it, the layout follows a predictable path pattern, and the README's own links reveal it: conference-publications/folders/years/2025/publications_neurips25/README.md, and the same shape for publications_icml25, publications_iclr25 and publications_kdd25.
The year segment is a plain four-digit folder. The leaf folder name concatenates the conference abbreviation and a two-digit year, so 2025 plus neurips becomes publications_neurips25. That convention is consistent enough that you can guess a path before you click it, which is a small but genuine convenience when you are scripting link checks or writing a scraper.
Older material is handled differently. The README routes 2017 and earlier to a single folder whose README uses anchor links, for example README.MD#neurips-2017 and README.MD#iclr-2016. Note the file extension there: README.MD in capitals, against README.md elsewhere. On a case-sensitive filesystem that distinction will break naive tooling that assumes lowercase. It is the kind of detail that only surfaces when something 404s.
Within each conference and year, the README says publications are "organised into topic-specific categories." The repository does not describe the category taxonomy, so the only way to learn it is to open a leaf README. Beyond papers, three side collections exist: Related Workshops, Surveys / Literature Reviews / Books, and Software / Libraries.
Reading a year folder without installing anything
There is nothing to install. The README gives no setup instructions because the deliverable is Markdown, and it publishes no commands of its own. What follows is the set of repository paths the README links to, written out so you know where to look rather than what to type.
The entry point for papers is this directory, which the README's conference links all resolve under:
conference-publications/folders/years/Below it, each year gets a four-digit folder, and each conference-year gets a leaf folder whose README holds the paper list. The README's own links give the shape directly:
conference-publications/folders/years/2025/publications_neurips25/README.md
conference-publications/folders/years/2025/publications_icml25/README.md
conference-publications/folders/years/2025/publications_iclr25/README.md
conference-publications/folders/years/2025/publications_kdd25/README.mdMaterial from 2017 and earlier is not split by year. The README points those conferences at a single folder and uses anchors inside its README, for example:
conference-publications/folders/years/2017_and_Earlier/README.MD#neurips-2017
conference-publications/folders/years/2017_and_Earlier/README.MD#iclr-2016The three side collections sit outside the year tree, under the same folders directory:
conference-publications/folders/workshops/README.md
conference-publications/folders/surveys/README.md
conference-publications/folders/software/README.mdOpen any of these files in an editor or on the web and you will see papers grouped under topic headings, each entry linking out to the paper itself. The README does not document the topic taxonomy, so the headings you find are the taxonomy.
The maintenance question, and what the last push date tells you
The last push to the default branch was on 2026-06-07. The repository is not archived. There are no releases, which is consistent with a project that ships Markdown rather than versioned artifacts.
A link index has an unusual maintenance profile. Its value decays not when the code rots but when the conferences it tracks stop being updated. The README already lists ICLR 2026, ICML 2026 and NeurIPS 2025 folders, so the index has been extended past the most recent completed cycles at the time of that push. Whether a given conference's folder is complete for a cycle is not something the README states, and there is no changelog to check against.
The second decay channel is link rot. Papers move between arXiv versions, workshop pages disappear, and project pages go offline. The repository has no visible link-checking infrastructure in its top-level layout, so a broken link is likely to persist until someone notices. If you build anything on top of this index, treat the URLs as a starting point and resolve them yourself.
A third consideration: there is no stated inclusion policy. The README does not say who curates the lists or what threshold a paper must clear to appear. That is a real gap if you intend to use absence from the index as evidence that a topic is not being worked on.
Where it stops being the right tool
The clearest failure mode is treating this as a source of implementations. It is not. The README points to a separate Software / Libraries page under conference-publications/folders/software/ for that purpose, which is a link list in its own right and not code hosted in this repository.
A second boundary is benchmarking. Nothing here reports accuracy, runtime or dataset results. If you need to compare two graph convolution variants on a standard split, this index will tell you the papers exist and nothing more.
A third is recency within a cycle. Conference proceedings appear in stages, and an index built from accepted-paper lists will lag the conference itself by however long the curator takes. For a topic moving quickly, a folder that looks thin may simply not have been updated yet.
Finally, the index is conference-scoped. Journal articles, technical reports and preprints that never landed at NeurIPS, ICML, ICLR or KDD are outside its frame. That is a defensible editorial choice, but it means a literature review built only on this repository will miss work that was published elsewhere. If your goal is exhaustive coverage, a citation database is the better instrument. If your goal is to know what the major venues accepted in a given year, this is faster than querying one by one.
Compared with a bibliographic database or a survey paper
The natural alternative is a citation database such as Semantic Scholar or Google Scholar, and the difference is one of framing rather than coverage. A database is query-driven and exhaustive: you supply keywords, it returns everything matching, ranked by its own signals. This repository is curation-driven and bounded: a human decided which papers count as graph-based deep learning and filed them by venue. You trade recall for a pre-filtered set that respects venue boundaries.
The second alternative is the survey paper. The README maintains a Surveys / Literature Reviews / Books page, which is an implicit acknowledgement that surveys and indexes serve different needs. A survey gives you the field's narrative, the taxonomy and the open problems, at the cost of being frozen at its publication date. This index gives you a continuously extended list with no narrative at all. Reading a survey tells you how to think about the area; reading this index tells you what appeared last year.
A third comparison is the awesome-list genre. Those usually mix papers, code, datasets and blog posts in one flat file. This repository separates papers from software and from surveys into distinct pages, which keeps each list readable. The cost is more clicking, and no single page that gives you the whole picture.
Licence and what MIT means for reuse
The repository is licensed under MIT, and the top level contains LICENSE.md. That covers the repository's own content: the README files, the folder structure, and the curation work itself.
It does not cover the papers being linked to. Each linked publication carries its own copyright, typically held by the authors or the publisher, and the licence terms vary between venues and between arXiv and the camera-ready version. MIT on this repository grants you nothing with respect to those documents.
In practice this means you can fork the index, reformat it, embed it in an internal tool or redistribute the link lists under the MIT terms. You cannot assume you may redistribute the papers themselves. If you are building a product that mirrors the index, check the licence on each linked artifact separately rather than inheriting this repository's terms. This is a description of how the licences are structured, not legal advice.
Editorial conclusion
Adopt it if you need a conference-scoped reading list for graph representation learning and you are willing to follow links out to arXiv or OpenReview for the actual papers. Do not adopt it if you want runnable implementations, benchmark numbers or a maintained library: the README points to a separate software and libraries page for that. Before relying on it, open the year folder for the conference you care about and confirm the papers you need are listed there, since coverage is organised per conference and year rather than as one searchable database.
Frequently asked questions
Are GCN and GNN the same?
No, and the repository's own topic list treats them as distinct labels: graph convolutional networks and graph neural networks both appear as topics. A graph convolutional network is one family of graph neural network, so the second term is the broader one. The index files papers under both labels, so a single search term will not surface everything.
Are graph neural networks still used?
The repository's structure is the evidence available here: it contains conference folders for NeurIPS, ICML, ICLR and KDD extending through 2026 in the README's own links, which indicates the venues are still accepting work in this area. The README does not make any claim about adoption in industry or about trends over time.
Is GNN better than CNN?
The repository does not answer this. It contains links to papers, and the README states no benchmark results, accuracy figures or comparisons between model families. Any claim about one architecture outperforming another would have to come from the individual papers, which you reach by following the links out of the index.
What is the difference between a graph and a network?
The repository does not define either term. It uses graph throughout its naming, in topics such as graph neural networks and graph representation learning, and its folder names follow the same convention. For a definitional answer you would need the surveys page the README links to under Surveys / Literature Reviews / Books.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/naganandy-graph-based-deep-learning-literature)