# safe-graph/graph-fraud-detection-papers: a curated paper index for graph and Transformer fraud detection > This repository is a reading list, not a library. It collects graph and Transformer papers on fraud, anomaly and outlier detection, plus a filterable dashboard and a separate local RAG chatbot, and it is aimed at researchers and engineers mapping the field before they build anything. **safe-graph/graph-fraud-detection-papers** — A curated list of Graph/Transformer-based fraud, anomaly, and outlier detection papers & resources - Repository: https://github.com/safe-graph/graph-fraud-detection-papers - Website: https://safe-graph.github.io/paper_dashboard/ - Stars: 1,897 · Forks: 299 - Language: Unknown - License: not declared - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/safe-graph-graph-fraud-detection-papers ## What this repository actually is, and who should read it This is an awesome-list style index, not software. The README opens by calling itself a curated list of Graph/Transformer-based papers and resources for fraud, anomaly, and outlier detection, and the repository layout confirms that scope: the top level contains README.md, and no source files, package manifest or build configuration appear alongside it. The value is in the tables, which are organised by year and venue and carry columns for Year, Title, Venue, Paper and Code. Entries span KDD, AAAI, ACL, EMNLP, NeurIPS, ACM MM and ACM CCS, alongside arXiv preprints and journals such as Expert Systems with Applications. The audience is narrow and specific. You are a graduate student or a research engineer who needs to know what has already been tried for fraud detection on graphs and with LLMs, and you want that map in one file rather than across a dozen survey PDFs. If your job is to ship a fraud model next sprint, this repository will not shorten that path, because nothing here executes. The topics list names fraud-detection, graph-neural-networks, anomaly-detection, outlier-detection, spam-detection, dataset and survey, which is an accurate description of an index rather than a toolkit. ## How the paper index is organised, and what the dashboard adds The table of contents splits the entries into six blocks: LLM and Transformer Papers, Deep Learning Graph Papers arranged by year from 2026 back to before 2020, Non-Deep-Learning Graph Papers since 2014, Toolbox, Dataset, Survey Paper, and Other Resource. The yearly grouping is the main navigation device, and each year heading links back to the top, which matters once you are deep in a long table. The second access path is the 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. That is the practical difference between this project and a plain markdown list: instead of scrolling a 2025 table to find text-attributed graph work, you filter. The README does not describe the dashboard's filter fields, its data source or how often it syncs with the README, so treat the dashboard as a convenience layer over the same list rather than an independent database. There is also a separate RAG-based chatbot repository, safe-graph/paper_chatbot, which the README describes as a local LLM chatbot built over 250 publicly accessible papers, with deployment instructions in that project's own README. That chatbot is a different codebase, not something you run from this repository. ## Getting the list onto your machine and searching it There is nothing to install. The README gives no install command, no package name and no dependency list, because the deliverable is a markdown file. The README states where to get the two companion tools instead: the dashboard lives at safe-graph.github.io/paper_dashboard, and the chatbot is a separate repository, safe-graph/paper_chatbot, whose README covers deployment. The realistic first use is to fetch the repository and read the tables, since the README does not document any command line workflow. The README does link to the repository's own GitHub page, and the badge row points readers at the dashboard and the chatbot repository for the two interactive views. Because the README describes no CLI, no search syntax and no local tooling, there is no command to reproduce here. Open README.md and use the table of contents to jump to the year or section you need, then follow the Paper link in each row. If you want filtering instead of scrolling, the README directs you to the dashboard, which is where the view, filter and search behaviour it advertises lives. ## Where the index breaks down as a working tool The Code column is inconsistent, and that is the biggest practical limitation. Many 2026 rows show a bare Link placeholder under Code, meaning the paper is indexed but no implementation is linked. Others do point at repositories, for example DGP under Xtra-Computing, UniDetect, PANTHER, OCR-APT and a text-attributed graph anomaly detection project. If your selection criterion is whether you can reproduce a result, you cannot use this list alone; you have to open each row and check. The second limitation is that the list is a snapshot of titles, not of findings. The README does not summarise any paper, does not record reported metrics, and does not flag which entries are preprints versus peer-reviewed. A 2026 arXiv row and a KDD 2026 row sit in the same table with the same treatment. The Venue column is the only signal, and it is not a quality ranking. Third, the dataset and toolbox sections are named in the table of contents but their contents are not visible in the README text, so it is not possible to confirm how many datasets are listed, what licences they carry, or whether they are downloadable. If you need data, verify each entry at its own source. Finally, the repository states no licence, which is worth resolving before you reuse the list wholesale in a paper or an internal report. ## Alternatives, and how they differ in approach The obvious alternative is a survey paper, and the repository itself points at one by keeping a Survey Paper section. The difference is direction of travel: a survey argues a taxonomy and tells you what the authors think matters, while this list refuses to rank and simply accumulates rows by year and venue. If you need a defensible framing of the field for a related-work section, a survey gives you the argument; the list gives you the raw citations to check it against. A second alternative is a general academic search engine or a preprint feed. Those are broader and updated continuously, but they are not scoped to fraud, anomaly and outlier detection on graphs, so you spend effort filtering out unrelated graph learning. The curated list has already made that cut, at the cost of depending on pull requests to stay current, which the README signals with a PRs-welcome badge. A third alternative sits inside the same project family: the paper_chatbot repository. Instead of scanning tables, you ask questions against a local retrieval-augmented LLM over 250 papers. That trades the list's transparency for convenience. A chatbot answer is harder to audit than a table row with a direct PDF link, and the README does not describe the chatbot's retrieval quality or coverage beyond the paper count. ## Maintenance, licence and the cost of keeping up The repository is not archived, and the last push was on 2026-06-29. The 2026 tables show entries from AAAI, KDD, ACL and arXiv, so the list is being extended rather than frozen. There are no releases, which is expected for a markdown index; nothing here is versioned for consumption by a dependency manager. The upgrade cost is therefore not a version bump, it is re-reading. Because entries are grouped by year, new work lands in new tables and the older years stay stable, so a fresh pull brings additions rather than rewrites of what you already cited. The maintenance burden falls on whoever curates: each new paper needs a row with a working paper link and, where one exists, a code link. Broken links are the failure mode you should expect over time, and the README does not describe any link-checking process. On licensing, the repository states no licence. That is a gap rather than a permission. Reusing a bibliography is usually low risk, but the linked papers, code repositories and datasets each carry their own terms, and none of those terms are recorded here. Check the licence of each asset you actually reuse, and treat the absence of a stated licence as something to resolve with the maintainers rather than as an implied grant. ## Conclusion Adopt it as a literature index if you are choosing a detection approach and need the LLM, Transformer and graph paper trail in one place, and expect to read the linked PDFs rather than run anything from this repository. Do not adopt it if you need a working detector, a maintained library, or a dataset you can download and train on today, because the README lists datasets and toolboxes by name without bundling data or code. Before relying on it, verify the licence, since none is stated in the repository, and check whether the papers you care about have a code link in the Code column, because many rows show only a paper link. ## FAQ ### What is graph-based fraud detection in the context of this repository? The repository indexes papers that model fraud, anomaly and outlier detection over graph structures, including graph neural networks and, in the newer tables, Transformer and LLM-based approaches. It is a reading list of that literature rather than an implementation of it. ### Does safe-graph/graph-fraud-detection-papers include code I can run? No. The top level contains only README.md, and the README gives no install steps. Some paper rows link to external code repositories in the Code column, but many rows show only a paper link, so availability varies entry by entry. ### How do I view or filter the papers in this list? The README points to an interactive dashboard at safe-graph.github.io/paper_dashboard, where the papers listed in the repo can be viewed, filtered and searched. The README does not document the dashboard's filter fields or how it syncs with the README. ### Is there a chatbot for these papers? Yes, but it is a separate project. The README describes a local RAG-based LLM chatbot at safe-graph/paper_chatbot built over 250 publicly accessible papers, and directs readers to that project's README for deployment. ## Sources - [Issues](https://github.com/safe-graph/graph-fraud-detection-papers/issues) - [Project website](https://safe-graph.github.io/paper_dashboard/) - [README](https://github.com/safe-graph/graph-fraud-detection-papers/blob/master/README.md) - [safe-graph/graph-fraud-detection-papers on GitHub](https://github.com/safe-graph/graph-fraud-detection-papers) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/safe-graph-graph-fraud-detection-papers