{
  "schema_version": 1,
  "title": "Repro: Optimal Fair Aggregation of Crowdsourced Noisy Labels using Demographic Parity Constraints",
  "emoji": "chart",
  "space_id": "snaykey/repro-fair-crowd-aggregation",
  "paper": {
    "openreview_id": "niQUth28zn"
  },
  "tags": [
    "icml2026-repro",
    "paper-niQUth28zn"
  ],
  "updated_at": "2026-07-27T00:00:00+00:00",
  "root": {
    "slug": "index",
    "title": "Repro: Optimal Fair Aggregation of Crowdsourced Noisy Labels using Demographic Parity Constraints",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1",
        "title": "The demographic parity gap of majority-vote aggregated labels is bounded above by an exponentially decaying term in the number of annotators R, of the form sum over groups a of E[e^{-R K_phi(a,X)}] (Proposition 3.2).",
        "file": "pages/claim-1/page.md",
        "children": []
      },
      {
        "slug": "claim-2",
        "title": "Under interpretable conditions on annotator skill, both Majority Vote and Bayesian aggregation are shown to achieve asymptotically unbiased fairness gaps relative to ground-truth labels as crowd size grows (Theorem 3.4).",
        "file": "pages/claim-2/page.md",
        "children": []
      },
      {
        "slug": "claim-3",
        "title": "For majority voting with R annotators, the aggregated fairness gap is bounded by epsilon(R) times the sum of individual annotators' fairness gaps, showing majority voting can amplify rather than average out bias (Proposition 3.6).",
        "file": "pages/claim-3/page.md",
        "children": []
      },
      {
        "slug": "claim-4",
        "title": "Convergence of the fairness gap to the ground-truth value requires a majority of annotators to have skill exceeding 0.5 plus a margin epsilon; without this, majority voting fails to converge (Section 3, following Proposition 3.6).",
        "file": "pages/claim-4/page.md",
        "children": []
      },
      {
        "slug": "claim-5",
        "title": "FairCrowd, a post-processing method, enforces a strict epsilon-fairness constraint on the output of any aggregation rule while provably bounding the resulting accuracy loss (Theorem 4.1).",
        "file": "pages/claim-5/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  }
}