logbook.json

{
  "schema_version": 1,
  "title": "Reproduction: Semi-knockoffs",
  "emoji": "\ud83c\udfad",
  "space_id": "ProCreations/repro-semi-knockoffs-audit",
  "paper": {
    "arxiv_id": "2601.23124",
    "openreview_id": "Xf9hJMGwDd"
  },
  "tags": [
    "icml2026-repro",
    "paper-Xf9hJMGwDd"
  ],
  "updated_at": "2026-07-28T05:00:00+00:00",
  "root": {
    "slug": "index",
    "title": "Reproduction: Semi-knockoffs",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-no-split-valid-pvalues",
        "title": "Claim 1: Semi-knockoffs avoids the train-test data split required by prior conditional-independence testing methods such as HRT while still yielding valid p-values, via nonparametric paired tests requiring only conditional expectations \u03bd_j and \u03c1_j rather than exact knockoff construction (Theorem 3.3, Section 3.1).",
        "file": "pages/claim-1-no-split-valid-pvalues/page.md",
        "children": []
      },
      {
        "slug": "claim-2-fdr-control",
        "title": "Claim 2: Theorem 3.4 establishes that the Semi-knockoffs procedure controls the false discovery rate at level q, i.e. FDR(S_SKO) \u2264 q (Theorem 3.4, Section 3.2).",
        "file": "pages/claim-2-fdr-control/page.md",
        "children": []
      },
      {
        "slug": "claim-3-optimization-stability",
        "title": "Claim 3: Theorem 4.1 shows that for null (non-relevant) features, regularized empirical risk minimizers trained with and without the feature remain close, with an \u2016\u03b8\u0303^j \u2212 \u03b8\u0302\u2016\u2082 \u2264 O_P(\u221a(log(1/\u03b4)/n)) bound, giving optimization stability guarantees for regularized models (Theorem 4.1, Section 4.2).",
        "file": "pages/claim-3-optimization-stability/page.md",
        "children": []
      },
      {
        "slug": "claim-4-double-robustness",
        "title": "Claim 4: Theorem 4.3 provides a double-robustness property: the loss difference between imputed feature distributions decays at a compound rate O_P(a_n b_n) even when both the predictive model and the sampler have estimation error (Theorem 4.3, Section 4.4).",
        "file": "pages/claim-4-double-robustness/page.md",
        "children": []
      },
      {
        "slug": "claim-5-power-vs-hrt-and-derandomisation",
        "title": "Claim 5: On simulated data with adjacent-feature support, Semi-knockoffs maintains type-I error control while achieving higher power than HRT, and derandomization with 5 permutations under masked correlation further increases power (Figure 4, Figure 5, Section 5.1).",
        "file": "pages/claim-5-power-vs-hrt-and-derandomisation/page.md",
        "children": []
      },
      {
        "slug": "claim-6-wisconsin-model-agnostic",
        "title": "Claim 6: On the Wisconsin Breast Cancer real dataset, Semi-knockoffs is applied across Random Forest, Neural Network, and Gradient Boosting models to demonstrate model-agnostic feature selection (Figure 6, Section 5.2).",
        "file": "pages/claim-6-wisconsin-model-agnostic/page.md",
        "children": []
      }
    ]
  }
}