CLAIMS.json

{
  "orid": "Xf9hJMGwDd",
  "arxiv_id": "2601.23124",
  "claims": [
    {
      "index": 1,
      "text": "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)."
    },
    {
      "index": 2,
      "text": "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)."
    },
    {
      "index": 3,
      "text": "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)."
    },
    {
      "index": 4,
      "text": "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)."
    },
    {
      "index": 5,
      "text": "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)."
    },
    {
      "index": 6,
      "text": "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)."
    }
  ]
}