safe-graph/graph-fraud-detection-papers: A Curated Reading List, Not a Detection System
A curated list of Graph/Transformer-based fraud, anomaly, and outlier detection papers & resources
At a glance
- What is it?
- This repository is an awesome-list of Graph and Transformer fraud, anomaly, and outlier detection papers, plus a dashboard and a separate RAG chatbot. It is a bibliography with navigation tooling, not installable detection code, and the README is the only specification you get.
- Who is it for?
- Adopt this list if you are a researcher or an engineer scoping graph or Transformer approaches to fraud and you need a year-sorted entry point; skip it if you need runnable detection code, because the repository is a bibliography and the README documents no installation of its own.
- Can I use it commercially?
- Not without permission. GitHub finds no licence file in the repository, and without a licence all rights are reserved by default: you may read the code but not reuse it. Check the README, or ask the authors, before using it.
- Is it still maintained?
- Yes. The repository last received commits 3 days ago.
- What is it written in?
- GitHub does not report a main language for this repository.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What the repository actually contains
The README describes a curated list of Graph and Transformer based papers and resources for fraud, anomaly, and outlier detection. The table of contents splits the material into LLM and Transformer papers, deep learning graph papers by year from 2026 back to before 2020, non-deep-learning graph papers since 2014, a toolbox section, a dataset section, a survey paper section, and an other resource section. Each table row carries a year, a title, a venue, a paper link, and in some rows a code link. That is the whole product: a structured index. There is no detection library here, no model checkpoint, no training script described in the README. The repository is an awesome-list, and the awesome.re badge in the header confirms the intent. If you arrived expecting to pip install something and score transactions, you have the wrong repository.
The dashboard and the RAG chatbot are separate artifacts
Two navigation aids sit on top of the list, and both live outside the repository itself. The first is an interactive dashboard at safe-graph.github.io/paper_dashboard, which the README says lets you view, filter, and search the papers listed in the repo. The second is a local RAG-based LLM chatbot maintained in a different repository, github.com/YingtongDou/paper_chatbot, which the README describes as covering 250 publicly accessible papers, with deployment instructions in that project's own README. Note the word local: the chatbot is something you deploy yourself, not a hosted service this repository runs. The dashboard is a static site under the safe-graph organisation. Neither is documented in detail here, so treat the dashboard as a search front end and the chatbot as a separate project with its own setup burden. The README does not specify the chatbot's model provider, vector store, or hardware requirements, so those are unknowns you would resolve in that other repository.
How the paper tables are organised, and why the year split matters
The LLM and Transformer table is sorted by year, and the deep learning graph papers are split into per-year sections from 2026 down to before 2020, with the non-deep-learning graph papers collected in a single section covering 2014 onward. That split is the useful editorial decision in the whole repository. Graph fraud detection has a long pre-neural lineage, and burying it in one section rather than deleting it keeps the classical work visible next to the newer Transformer entries. The LLM table shows where the field has moved: entries such as TransactionGPT at KDD 2026, TREASURE at KDD 2026, PRAGMA described as a Revolut foundation model, and DGP, a dual-granularity prompting framework at AAAI 2026, are all transaction or fraud foundation model work. A reader can scan that table and see that the 2025 to 2026 cohort is dominated by language-model pretraining on financial sequences and by LLM plus GNN hybrids. That is a real signal, and it comes from the table structure rather than from any prose the repository writes.
What you can and cannot run from this repository
The README gives no install command for the list itself, because there is nothing to install. The one deployment instruction it does give points elsewhere: to deploy the chatbot, refer to the project README in the paper_chatbot repository. The dashboard is reached by URL, not by a local command. So the honest answer to how you get this running is that you do not run the list. You either open the dashboard in a browser, clone the repository to read the tables as markdown, or go to the separate chatbot repository and follow its instructions. The README does not state a licence for this repository, and the repository metadata supplied does not carry one either. If you intend to reuse the tables in a derived work, that gap is the first thing to resolve, and it is a question for the maintainers rather than something the README answers.
The Code column is a pointer, not a guarantee
Many rows in the LLM table have a Code cell containing a link, and several of those links resolve to real repositories, for example the DGP entry pointing at github.com/Xtra-Computing/DGP, UniDetect at github.com/msy0513/UniDetect, OCR-APT at github.com/CoDS-GCS/OCR-APT, and TAG_AD at github.com/Flanders1914/TAG_AD. Others carry the literal word Link with no target, or an anonymous.4open.science URL that is a review artifact rather than a permanent home. The list does not record whether a given code link is maintained, what framework it targets, or whether it reproduces the paper's numbers. A reader who picks a paper because it has a Code cell may still find a repository that last saw a commit years ago. That is the main failure mode of any curated list, and this one does not mitigate it with status annotations. Verify each code link individually before you build on it.
Where a maintained library beats a reading list
The toolbox section is the closest this repository comes to pointing at software, but it is still a set of pointers. If your goal is to run graph-based fraud detection on your own data rather than to survey the literature, the practical difference is stark: a library such as PyTorch Geometric or DGL gives you message-passing primitives, dataset loaders, and a documented API surface, while this repository gives you citations to papers that use those primitives. The two are complementary but not substitutes. A list cannot tell you whether a model trains on your graph schema, what memory it needs, or how it behaves under class imbalance. Its value is in narrowing the search space before you commit engineering time. If you already know which architecture family you want, the list adds little; if you are deciding between a GNN baseline and an LLM-driven agent, the year-sorted tables give you the candidate set quickly.
Maintenance, freshness, and the cost of following it
The repository is not archived and the last push recorded is 2026-06-29, so it is being touched. No releases have been retrieved, which is expected for a list that ships as markdown rather than versioned artifacts. The maintenance model is pull requests: the README carries a PRs-welcome badge and links to makeapullrequest.com. That means quality depends on contributors keeping entries accurate, and there is no stated review cadence or inclusion policy in the material. The upgrade cost is therefore near zero in the software sense, because there is no dependency to bump, but non-zero in the editorial sense, because you have to re-check the tables when you rely on them. The licence is not stated, so redistribution terms are unclear. For an internal reading list that is tolerable; for anything you publish or ship, resolve the licence first.
Who should bookmark this, and who should keep looking
This is a resource for researchers, graduate students, and engineers doing a literature scan before choosing an approach to graph or Transformer based fraud detection. It is also useful for someone writing a survey or a proposal who needs a year-organised entry point and a filterable dashboard. It is the wrong tool if you need a detection system today, if you need benchmark numbers to compare methods, or if you need a maintained library with an API. Those needs point to the individual paper repositories and to graph learning frameworks, not to this index. The concrete next step for an adopter is to open the dashboard, filter to the years and venues that match your problem, and then check the Code column of each shortlisted paper against its actual repository activity, because the list itself will not tell you whether the code still runs.
Editorial conclusion
Adopt this list if you are a researcher or an engineer scoping graph or Transformer approaches to fraud and you need a year-sorted entry point; skip it if you need runnable detection code, because the repository is a bibliography and the README documents no installation of its own. Before relying on it, verify the licence and maintenance status, since the repository metadata does not state a licence and no releases have been retrieved, and confirm that the linked paper code repositories are still reachable rather than assuming the table's Code column resolves.
Community notes