Datasets:

Ngseo
/

machine-unlearning-papers

license: other
license_name: metadata-compilation
license_details: >-
  Bibliographic metadata compiled from DBLP (CC0), Semantic Scholar (ODC-BY),
  and OpenReview. PDFs in pdfs/ retain their original publisher/author
  copyrights and are stored here for personal research use only — do not make
  this repository public with the PDFs included.
language:
  - en
task_categories:
  - text-classification
tags:
  - machine-unlearning
  - literature-survey
  - bibliography
pretty_name: Machine Unlearning Papers at Top-Tier Venues (2016-2026)
size_categories:
  - n<1K

Machine Unlearning Papers at Top-Tier Venues (2016–2026)

Complete crawl of machine-unlearning papers from 9 top-tier venues: IEEE S&P, USENIX Security, ACM CCS, NDSS (security) and NeurIPS, ICML, ICLR, AAAI, IJCAI (AI/ML), years 2016–2026.

501 core papers (+ 14 adjacent-field papers flagged needs_review)
454 PDFs under pdfs/ (<year>_<venue>_<slug>.pdf)
Built 2026-08 with a multi-stage pipeline (DBLP + Semantic Scholar + OpenReview) followed by a multi-agent coverage audit against official proceedings; 539 false positives were removed (workshop tracks, student abstracts, ICMLA/ICMLC mismatches, continual-learning noise) and 67 audit-discovered missing papers were added after per-paper verification.
2026 is partial: NeurIPS 2026 and CCS 2026 had not taken place at crawl time.

Files

file	description
`papers.jsonl` / `papers.csv`	final list; one row per paper
`meta/excluded.jsonl`	removed records with `excluded_reason` (for recovery/audit)
`meta/triage.json`	human/agent verdict overrides applied during merging
`meta/pdf_report.json`, `meta/pdf_missing_report.json`	PDF acquisition log; 47 core papers lack PDFs (OpenReview-only, no arXiv preprint)
`meta/wf_result.json`	raw multi-agent audit output
`pdfs/`	collected PDFs (arXiv preprints or official open-access versions)

Record fields

title, authors, year, venue, venue_key, field (security/ai), status (core / needs_review), abstract, doi, arxiv_id, pdf_url, pdf_candidates, url, citations (Semantic Scholar, 2026-08), sources (dblp/s2/openreview/manual).

Scope

Core = exact/approximate unlearning, certified data removal, federated/graph/recommendation unlearning, LLM knowledge unlearning, concept erasure in generative models/representations, unlearning evaluation/verification/auditing, attacks on unlearning, right-to-be-forgotten deletion from trained models. needs_review = adjacent literatures that remove training-data influence but position themselves elsewhere (backdoor purification, model repair, deletion-robust optimization).

Excluded: catastrophic forgetting / continual learning, unlearnable examples (availability poisoning), logic-based forgetting (KR), database deletion.