Pith. sign in

REVIEW 5 major objections 6 minor 35 references

Knockoff-Guided Compressive Sensing: A Statistical Machine Learning Framework for Support-Assured Signal Recovery

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Knockoffs give sparse-signal recovery a provable false-discovery cap

desk verdict The paper's central knockoff construction degenerates in the compressive sensing regime it targets (m < n forces s = 0, so the knockoff equals the original and all statistics vanish), and the experiments never run the proposed algorithm. read the letter →

arxiv 2505.24727 v1 pith:ZHJVPMM3 submitted 2025-05-30 stat.ML cs.LGeess.SP

Xiaochen Zhang , Haoyi Xiong This is my paper
classification stat.MLcs.LGeess.SP MSC 62F0362J07
keywords compressivesensingknockofffilterfalsediscoveryratesupportrecoverysparsesignalreconstructionFDRcontrolvariableselectionunderdeterminedlinearsystems
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes KnockoffCS, a compressive sensing method that splits sparse-signal recovery into two stages: support identification with false discovery rate (FDR) control, then least-squares reconstruction on the selected support. The central claim is that by constructing a knockoff measurement matrix $\tilde{A}$ alongside the true matrix $A$ and comparing the per-coordinate correlations $|A^\top y|_j$ with $|\tilde{A}^\top y|_j$, one can select the nonzero entries of $x$ while guaranteeing that the expected proportion of false discoveries stays below a preset level $q$. The paper further claims this FDR control comes with reconstruction error of order $\sigma\sqrt{s\log n / m}$, a rate comparable to $\ell_1$-based compressive sensing but obtained under weaker distributional assumptions. If the guarantees hold, users of compressed sensing would gain a principled, interpretable statistical bound on which detected support entries are real, alongside the usual recovery accuracy.

What carries the argument

The central object is the knockoff measurement matrix $\tilde{A}$, constructed so that $\tilde{A}^\top \tilde{A} = A^\top A$ and $\tilde{A}^\top A = A^\top A - \operatorname{diag}(s)$ with $s = \min\{1, 2\lambda_{\min}(A^\top A)\}$. The per-coordinate statistics $W_j = |[A^\top y]_j| - |[\tilde{A}^\top y]_j|$ convert the knockoff comparison into a signed score: positive values suggest a true support coordinate, while negative values serve as negative controls. The threshold rule in equation (9) estimates the false discovery proportion from the count of large negative versus large positive statistics, and the support estimate $\hat{S}$ is the set of coordinates above the chosen threshold. The proof connects FDR control to support recovery through Assumption A1.1(b), which postulates a $\delta$-gap between $|\alpha_j| = |[A^\top A x]_j|$ and $|\beta_j| = |[\tilde{A}^\top A x]_j|$ on the true support; this gap is what lets the tail bound in Theorem 2 turn the FDR property into a probability that the true support is included.

What would settle it

In the underdetermined case $m < n$, set $A \in \mathbb{R}^{m \times n}$ with rank $m$ so that $A^\top A$ is singular; Algorithm 1 then computes $\lambda_{\min}(A^\top A) = 0$, $s = 0$, $\tilde{A} = A$, and every $W_j = |[A^\top y]_j| - |[\tilde{A}^\top y]_j| = 0$. With all statistics zero no positive threshold $T$ exists and the estimated FDR and support-inclusion bounds of Proposition 1 have no content; running the procedure on such a matrix and checking whether the claimed FDR guarantee and recovery probability hold would settle the central claim.

Watch

Extended reading notes

Core claim

The core discovery the paper aims to establish is that the knockoff filter, a regression tool for FDR-controlled variable selection, can be transplanted onto the measurement matrix of a compressive sensing problem. The paper constructs $\tilde{A} = A(I - C)$ from $A$ and defines statistics $W_j = |[A^\top y]_j| - |[\tilde{A}^\top y]_j|$, so that coordinates whose original correlation exceeds their knockoff's correlation can be tested as a group. A threshold $T$ chosen by the rule in equation (9) yields the support estimate $\hat{S} = \{j : W_j \ge T\}$, and the theoretical section (Proposition 1 and Theorems 2–5) states that this procedure controls FDR at level $q$, achieves support-recovery probability at least $1 - 4s\exp(-(\delta-T)^2/(8\sigma^2 m))$, and bounds reconstruction error by $\frac{\sqrt{C}\,\sigma(\sqrt{s} + \sqrt{|S_f|})}{\kappa_{\min}^2 - \gamma}\sqrt{\frac{\log n}{m}}$, refining to $O(\sigma\sqrt{s\log n/m})$ when $q = O(s/m)$. The experiments claim these guarantees translate into F1-score gains up to $3.9\times$ over LASSO-based and OMP-based compressive sensing on synthetic data, and improved regression and classification accuracy on eight real-world benchmark datasets.

Load-bearing premise

The entire argument rests on the claim that Algorithm 1 produces a knockoff matrix $\tilde{A}$ that is genuinely different from $A$ and leaves a positive gap $\delta$ between the signal's alignment with the original and knockoff measurements on the true support; if $A^\top A$ is singular the construction sets $\tilde{A}=A$ and every statistic $W_j$ is zero, and the paper does not establish such a gap in general.

Editorial extensions

If this is right

  • If the construction is valid, compressive sensing users can prescribe a target FDR $q$ and obtain a support estimate whose expected fraction of false positives is at most $q$, without relying on restricted-isometry-type worst-case bounds.
  • The bound $P(S \subseteq \hat{S}) \ge 1 - 4s\exp(-(\delta-T)^2/(8\sigma^2 m))$ implies that at high signal-to-noise ratio the true nonzeros are almost surely included in the selected support.
  • The stated reconstruction rate $O(\sigma\sqrt{s\log n/m})$ matches the classical compressive sensing rate, suggesting that decoupling FDR-controlled selection from least-squares refitting does not pay a rate penalty.
  • The reported F1-score improvements of up to $3.9\times$ over LASSO and OMP on simulated data, and top-two performance in 71.4% of real-world regression and classification model–dataset pairs, indicate that the FDR-controlled support yields better downstream predictive features.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not examine the underdetermined regime $m < n$ where $A^\top A$ is singular: then $\lambda_{\min} = 0$, $s = 0$, $\tilde{A} = A$, and every $W_j$ is identically zero, so a natural extension is to replace $\lambda_{\min}$ by a restricted eigenvalue and check whether the FDR claim survives numerically.
  • A practical diagnostic the paper leaves implicit: plotting the distribution of $W_j$ for a mass near zero would reveal whether the $\delta$-gap assumption holds, since a spike at zero signals that the knockoff carries no information.
  • The recipe generalizes beyond compressive sensing: any sparse inverse problem with a known forward operator could equip its operator with a knockoff copy to attach finite-sample FDR guarantees to support identification.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The manuscript proposes KnockoffCS, an algorithm that augments a compressive sensing measurement matrix A with a knockoff matrix \tilde A constructed in Algorithm 1, forms statistics W_j = |[A^T y]_j| - |[\tilde A^T y]_j|, selects a support set by an FDR threshold, and then estimates the signal by least squares or ridge regression on the selected support. The main theoretical claims are finite-sample FDR control for support recovery (Theorem 2), a support-recovery probability bound, and reconstruction error bounds of order O(\sigma \sqrt{s \log n / m}) under restricted eigenvalue and coherence assumptions (Theorems 3-5). The paper also reports simulation studies and downstream prediction experiments on eight real-world datasets.

Significance. If the FDR-control and recovery guarantees were valid, the framework would be a meaningful contribution: it would add finite-sample false-discovery control to underdetermined compressive sensing support recovery and provide a principled alternative to Lasso-based reconstruction. The paper also makes some reproducibility efforts by linking to code. However, the central construction degenerates in the m<n regime that the paper explicitly targets, the FDR proof is imported from the existing knockoff literature without verifying the required sign-flip property for the proposed statistic, and the experiments do not implement Algorithm 1. As a result, the current manuscript does not establish its main claims.

major comments (5)
  1. [Algorithm 1, §3.2] In the underdetermined compressive sensing setting m<n (Eq. (1), §2.1), the matrix \Sigma = A^T A has rank at most m, so \lambda_{\min}(\Sigma)=0. Line 2 of Algorithm 1 then sets s=0, and since \Sigma^{-1/2} is not defined for singular \Sigma (or, with a pseudoinverse, gives C=0), \tilde A = A(I-C)=A. Consequently every statistic W_j in Eq. (8) is identically zero, the threshold condition (9) is satisfied trivially for every t>0, and the support set from Eq. (10) is either empty or the full index set. The FDR bound (23), the support-recovery probability (24), and Assumption A1.1(b), which requires |\alpha_j| \ge |\beta_j|+\delta while \beta_j=\alpha_j when \tilde A=A, are therefore vacuous or false in the intended regime.
  2. [Theorem 2, §4.1] The proof of the FDR bound (23) is deferred to [22] without establishing that the statistic W_j of Eq. (8) satisfies the required sign-flip or symmetry property for null variables; this property is not automatic for statistics of the form |[A^T y]_j| - |[\tilde A^T y]_j|. In addition, the probability direction in Eq. (32) is reversed: from W_j \ge \delta - (|\gamma_j|+|\tilde\gamma_j|), the event |\gamma_j|+|\tilde\gamma_j| \le \delta-T implies W_j \ge T, so P(W_j \ge T) \ge P(|\gamma_j|+|\tilde\gamma_j| \le \delta-T), not the reverse. The argument needs a correct lower-tail bound on W_j and a proof of the knockoff sign-flip property.
  3. [Theorem 3, Assumption 2] The reconstruction guarantee in Eq. (42) depends on |S_f| = |\hat S \setminus S|, and Assumptions A2.2-A2.3 impose conditions on A_{S_f}, i.e., on the method's own false-discovery set. A probability statement of the form 'with probability at least 1 - 4s exp(...) - exp(-C|S_f|/8 log n)' cannot be verified before seeing \hat S; the bound is therefore conditional on the very quantity the method is supposed to control. This makes the reconstruction guarantee in Theorem 3 not a valid a priori error bound as stated, and the issue propagates to Theorems 4 and 5.
  4. [Theorem 4] In the proof of Theorem 4, Markov's inequality gives P(|S_f|/|\hat S| \ge q/\eta) \le \eta, so the useful event is |S_f| \le |\hat S| q/\eta. Equation (86) instead uses \sqrt{|\hat S| q} (dropping the 1/\eta factor), and Eq. (87) contains an unexplained factor \sqrt{m q \log n / m}. As written, the chain from (85) to (88) does not follow.
  5. [§5] The experimental section does not evaluate Algorithm 1. Section 5.1.1 states that 'The Knockoff-based selection in KnockoffCS is implemented using the knockpy library, adopting a model-X knockoff construction' with Lasso coefficient differences, and Section 5.2 selects 'the top 1% of features ranked by their knockoff statistics'. Neither matches the fixed-X construction of Algorithm 1 with W_j of Eq. (8) and the FDR threshold of Eq. (9). The reported F1-score, FDR, power, and reconstruction gains therefore do not test the proposed method or its guarantees.
minor comments (6)
  1. [§1, §2.1] Notation is inconsistent: the introduction uses x \in R^p and A \in R^{n \times p}, while §2.1 and Algorithm 1 use x \in R^n and A \in R^{m \times n}; this makes it difficult to track which dimension is the number of measurements.
  2. [Algorithm 1] Line 2 writes s = min{1, 2\lambda_{\min}(\Sigma)}1, mixing a scalar and a vector, and line 3 uses \Sigma^{-1/2} without defining the matrix square-root convention for singular \Sigma; Eq. (7) uses a diagonal matrix S while the algorithm uses a vector s.
  3. [Eq. (9)] The 'adjust' operator in Eq. (9) is never defined; if it is meant to be the Benjamini-Hochberg correction, the selection rule (9)-(10) should specify exactly how the adjusted threshold is computed.
  4. [Theorem 2, Eq. (24)] The support-recovery probability bound is only meaningful when T < \delta, but no condition ensuring this is stated; T is data-dependent while \delta is a fixed constant from Assumption A1.1(b).
  5. [References, §1] Some references are mismatched with the claims they support: for example, the group-sparse and high-dimensional extensions cited in §1 do not correspond to the listed references [13]-[15], and [12] is cited both as the knockoff filter and as a group-sparse/multitask paper.
  6. [§5.1.1] The statement that 'The knockoff selection ratio is fixed at 0.1 for all simulations' is never defined or connected to Algorithm 1 or to the FDR threshold in Eq. (9).

Circularity Check

2 steps flagged · score 6.0 of 10

The support-recovery guarantee is assumed through Assumption A1.1(b) rather than produced by Algorithm 1, and the reconstruction bound is conditioned on the method's own false-discovery output.

  1. self definitional [Section 4.1, Assumption 1(A1.1)(b), Eqs. (21)-(22); Theorem 2, Eq. (24); Algorithm 1, lines 2-4]
    "Set s ← min{1, 2λmin(Σ)}1 ... Generate à ← A(I − C) ... The knockoff matrix à ∈ R^{m×n} is constructed such that ∃δ > 0, for every j in the true support set S, the following condition holds: |αj| ≥ |βj| + δ, where αj = [A⊤Ax]j and βj = [Ã⊤Ax]j."

    In the paper's own target regime m ≪ n (Section 2.1), rank(Σ)=rank(A⊤A)≤m<n, so λmin(Σ)=0. Algorithm 1 then forces s=0, C=0, and Ã=A(I−C)=A, giving βj=[A⊤Ax]j=αj for every j. The δ-separation in Assumption A1.1(b) is therefore impossible by construction, and it is exactly the separation needed to conclude Wj≥T for true-support coordinates. The support-recovery bound (24) is not derived from the algorithm; it is an unpacking of an assumption that already contains the desired conclusion, and in the compressive-sensing setting that assumption is contradicted by the paper's own construction.

  2. other [Section 4.2, Assumption 2 (A2.2)-(A2.4); Theorem 3, Eq. (42)]
    "Mutual Coherence Condition: ... where Sc = Ŝ ∩ S is the correctly identified support set and Sf = Ŝ \ S is the false discovery support set. Orthogonality Condition: The columns of AS and ASf are orthogonal. ... ∥x̂ − x∥2 ≤ √Cσ(√s + √|Sf|)/(κ²min − γ) √(log n/m)."

    The reconstruction guarantee is stated under hypotheses that are properties of the method's own output: Sf = Ŝ \ S is determined by the threshold step, and the theorem's error bound contains |Sf| on the right-hand side. The theorem applies only when the false-discovery set happens to satisfy coherence and orthogonality with the true support, and its bound automatically worsens as |Sf| grows. Thus the claimed O(σ√(s log n/m)) rate is not a forward guarantee that KnockoffCS achieves the rate; it is a self-referential statement about favorable outputs, with the output's own false-discovery size appearing in both the assumptions and the bound.

full rationale

KnockoffCS does not fit parameters and does not depend on the authors' self-citations, so the circularity is not of the tuning-parameter or self-citation types. However, the central derivation chain reduces at two load-bearing points. First, the support-recovery guarantee is not produced by Algorithm 1: it is assumed through the δ-separation in Assumption A1.1(b), which is precisely the property that true-support knockoff statistics dominate the knockoff statistics, and which Algorithm 1 cannot deliver in the m≪n regime because it yields Ã=A and Wj≡0. Second, the reconstruction error bound is conditioned on Sf=Ŝ\S, the method's own false-discovery output, and the bound contains |Sf|; so the theorem proves a bound for outputs satisfying conditions phrased in terms of themselves, rather than an independent guarantee that the algorithm attains the stated rate. These are genuine reductions of the claimed predictions to self-referential inputs, though the absence of fitted inputs and the use of external knockoff theory (rather than author self-citation) keep the score below 8.

Assumptions & free parameters 1 free parameters · 6 assumptions · 1 invented entities

The framework depends on a large set of unverified premises: a valid fixed-X knockoff construction, a delta-separation between signal and knockoff interactions, Gaussian noise, restricted eigenvalue and coherence conditions, and orthogonality involving the method's own false-discovery set. The construction itself degenerates for m<n. No constants are fitted to data in the theory, so circularity burden comes mainly from output-dependent assumptions.

free parameters (1)
  • Target FDR level q = 0.1 in experiments; set to c*s/m in Theorem 4
    User-chosen threshold that sets T and controls the trade-off; Theorem 4 additionally requires knowing sparsity s to choose q.
assumptions (6)
  • ad hoc to paper Fixed-X knockoff construction satisfies Eq. (7)
    Section 3.2 claims A~ = A(I - C) satisfies the covariance identity, but for m<n the choice s = min{1, 2*lambda_min(A^T A)} gives s=0, A~=A, which cannot provide knockoff control.
  • ad hoc to paper delta-separation in Assumption A1.1(b): |alpha_j| >= |beta_j| + delta for all j in S
    Used in Theorem 2 proof to ensure the true support passes the threshold; not guaranteed by the construction and not verified in simulations.
  • domain assumption Gaussian noise w ~ N(0, sigma^2/m I_m)
    Theorem 2 and Lemma 1 rely on Gaussian tail bounds and chi-square moment generating functions; the abstract claims model-free benefit but the theory is noise-specific.
  • domain assumption Restricted eigenvalue condition (A2.1)
    Invoked in Theorem 3; no verification for the block-correlated A used in experiments.
  • ad hoc to paper Mutual coherence and orthogonality of A_S and A_{S_f} (A2.2, A2.3)
    These bounds are stated in terms of S_f = hat S \ S, the method's own false discoveries, so they cannot be checked before the method runs.
  • domain assumption Estimated support size bound |hat S| <= m (A2.4)
    Assumed in Theorem 3; plausible but not guaranteed by the FDR threshold.
invented entities (1)
  • Knockoff measurement matrix A~ = A(I - C)
    purpose: Negative-control measurement matrix intended to enable FDR-controlled support testing.
    Introduced in Section 3.2; no external evidence. The construction degenerates to A~=A when m<n, so the proposed entity does not function as intended.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Knockoff-Guided Compressive Sensing: A Statistical Machine Learning Framework for Support-Assured Signal Recovery." pith.science (2026). https://pith.science/paper/ZHJVPMM3

@misc{pith2026250524727,
  author       = {Pith},
  title        = {Pith review of: Knockoff-Guided Compressive Sensing: A Statistical Machine Learning Framework for Support-Assured Signal Recovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZHJVPMM3}},
  note         = {Machine review of arXiv:2505.24727}
}
abstract

This paper introduces a novel Knockoff-guided compressive sensing framework, referred to as \TheName{}, which enhances signal recovery by leveraging precise false discovery rate (FDR) control during the support identification phase. Unlike LASSO, which jointly performs support selection and signal estimation without explicit error control, our method guarantees FDR control in finite samples, enabling more reliable identification of the true signal support. By separating and controlling the support recovery process through statistical Knockoff filters, our framework achieves more accurate signal reconstruction, especially in challenging scenarios where traditional methods fail. We establish theoretical guarantees demonstrating how FDR control directly ensures recovery performance under weaker conditions than traditional $\ell_1$-based compressive sensing methods, while maintaining accurate signal reconstruction. Extensive numerical experiments demonstrate that our proposed Knockoff-based method consistently outperforms LASSO-based and other state-of-the-art compressive sensing techniques. In simulation studies, our method improves F1-score by up to 3.9x over baseline methods, attributed to principled false discovery rate (FDR) control and enhanced support recovery. The method also consistently yields lower reconstruction and relative errors. We further validate the framework on real-world datasets, where it achieves top downstream predictive performance across both regression and classification tasks, often narrowing or even surpassing the performance gap relative to uncompressed signals. These results establish \TheName{} as a robust and practical alternative to existing approaches, offering both theoretical guarantees and strong empirical performance through statistically grounded support selection.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 28 canonical work pages

  1. [22]

    The Annals of Statistics 43(5), 2055–2085 (2015) https://doi.org/10.1214/ 32 15-AOS1337

    Barber, R.F., Cand` es, E.J.: Controlling the false discovery rate via knock- offs. The Annals of Statistics 43(5), 2055–2085 (2015) https://doi.org/10.1214/ 32 15-AOS1337

    work page 2015

  2. [1]

    IEEE Transactions on information theory 52(2), 489–509 (2006)

    Cand` es, E.J., Romberg, J., Tao, T.: Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information. IEEE Transactions on information theory 52(2), 489–509 (2006)

    work page 2006

  3. [2]

    IEEE signal processing magazine 25(2), 21–30 (2008)

    Cand` es, E.J., Wakin, M.B.: An introduction to compressive sampling. IEEE signal processing magazine 25(2), 21–30 (2008)

    work page 2008

  4. [3]

    IEEE Transactions on Signal Processing (2008)

    Baron, D., Sarvotham, S., Baraniuk, R.: Bayesian compressive sensing via belief propagation. IEEE Transactions on Signal Processing (2008)

    work page 2008

  5. [4]

    Khosravy, M., Dey, N., Duque, C.A.: Compressive Sensing in Healthcare, pp. 1–3. Academic Press, ??? (2020)

    work page 2020

  6. [5]

    IEEE Transactions on signal processing 59(9), 4053–4085 (2011)

    Duarte, M.F., Eldar, Y.C.: Structured compressed sensing: From theory to applications. IEEE Transactions on signal processing 59(9), 4053–4085 (2011)

    work page 2011

  7. [6]

    arXiv preprint (2009)

    Hormati, A., Karbasi, A., Mohajer, S., Vetterli, M.: An estimation theoretic approach for sparsity pattern recovery in the noisy setting. arXiv preprint (2009)

    work page 2009

  8. [7]

    In: European Symposium on Security and Privacy (2020) 31

    Kerkouche, R., ´Acs, G., Castelluccia, C., Genev` es, P.: Compression boosts dif- ferentially private federated learning. In: European Symposium on Security and Privacy (2020) 31

    work page 2020

Show all 35 references
  1. [8]

    Scandinavian Journal of Statistics (2022)

    Tardivel, P.J., Bogdan, M.: On the sign recovery by lasso, thresholded lasso and thresholded basis pursuit denoising. Scandinavian Journal of Statistics (2022)

    2022

  2. [9]

    IEEE Transactions on Signal and Information Processing over Networks (2016)

    Wimalajeewa, T., Varshney, P.: Sparse signal detection with compressive mea- surements via partial support set estimation. IEEE Transactions on Signal and Information Processing over Networks (2016)

    2016

  3. [10]

    IEEE Transactions on Computational Imaging (2022)

    Niu, J.: Robust deep compressive sensing with recurrent-residual structural constraints. IEEE Transactions on Computational Imaging (2022)

    2022

  4. [11]

    Annals of Statistics (2016)

    Barber, R., Cand` es, E.: A knockoff filter for high-dimensional selective inference. Annals of Statistics (2016)

    2016

  5. [12]

    In: International Conference on Machine Learning (2016)

    Dai, R., Barber, R.: The knockoff filter for fdr control in group-sparse and multitask regression. In: International Conference on Machine Learning (2016)

    2016

  6. [13]

    Ren, Z., Barber, R.: Derandomized knockoffs: leveraging e-values for false discovery rate control (2022)

    2022

  7. [14]

    Journal of the Royal Statistical Society Series B: Statistical Methodology (2021)

    Cao, Y., Sun, X., Yao, Y.: Controlling the false discovery rate in transforma- tional sparsity: Split knockoffs. Journal of the Royal Statistical Society Series B: Statistical Methodology (2021)

    2021

  8. [15]

    Machkour, J., Muma, M., Palomar, D.: The terminating-random experiments selector: Fast high-dimensional variable selection with false discovery rate control (2021)

    2021

  9. [16]

    arXiv preprint (2015)

    Cand` es, E., Su, W.J.: Slope is adaptive to unknown sparsity and asymptotically minimax. arXiv preprint (2015)

    2015

  10. [17]

    In: IEEE International Conference on Acoustics, Speech, and Signal Processing (2024)

    Machkour, J., Breloy, A., Muma, M., Palomar, D., Pascal, F.: Sparse pca with false discovery rate controlled variable selection. In: IEEE International Conference on Acoustics, Speech, and Signal Processing (2024)

    2024

  11. [18]

    Journal of the Royal Statistical Society Series B: Statistical Methodology 58(1), 267–288 (1996)

    Tibshirani, R.: Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology 58(1), 267–288 (1996)

    1996

  12. [19]

    Journal of the American Statistical Association (2019)

    Xing, X., Zhao, Z., Liu, J.S.: Controlling false discovery rate using gaussian mirrors. Journal of the American Statistical Association (2019)

    2019

  13. [20]

    Abramovich, F., Benjamini, Y., Donoho, D., Johnstone, I.: Adapting to unknown sparsity by controlling the false discovery rate (2005)

    2005

  14. [21]

    Emery, K., Keich, U.: Controlling the fdr in variable selection via multiple knockoffs (2019)

    2019

  15. [23]

    Journal of the Royal Statistical Society: Series B (Statistical Methodology) 80(3), 551–577 (2018) https://doi

    Cand` es, E., Fan, Y., Janson, L., Lv, J.: Panning for gold: Model-X knockoffs for high-dimensional controlled variable selection. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 80(3), 551–577 (2018) https://doi. org/10.1111/rssb.12265

    2018 doi

  16. [24]

    Journal of the Royal statistical society: series B (Methodological) 57(1), 289–300 (1995)

    Benjamini, Y., Hochberg, Y.: Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal statistical society: series B (Methodological) 57(1), 289–300 (1995)

    1995

  17. [25]

    Wiley Interdisciplinary Reviews: Computa- tional Statistics 1(1), 93–100 (2009)

    McDonald, G.C.: Ridge regression. Wiley Interdisciplinary Reviews: Computa- tional Statistics 1(1), 93–100 (2009)

    2009

  18. [26]

    PhysioNet (2023)

    Wang, Q., Zhou, Y., Ghassemi, P., Chenna, D., Chen, M., Casamento, J., Pfefer, J., McBride, D.: Facial and oral temperature data from a large set of human subject volunteers (version 1.0.0). PhysioNet (2023)

    2023

  19. [27]

    https://ssd.jpl.nasa.gov/ sbdb query.cgi

    Technology, J.P.L.: Mars Asteroid Observation Dataset. https://ssd.jpl.nasa.gov/ sbdb query.cgi. Accessed: 2025-05-10 (2024)

    2024

  20. [28]

    In: Proceedings of the 12th International Conference on Music Information Retrieval (ISMIR 2011) (2011)

    Bertin-Mahieux, T., Ellis, D.P.W., Whitman, B., Lamere, P.: The million song dataset. In: Proceedings of the 12th International Conference on Music Information Retrieval (ISMIR 2011) (2011)

    2011

  21. [29]

    IEEE Access 13, 35750–35779 (2025) https://doi.org/10.1109/ACCESS.2025.3543475

    Mathur, P., Shaikh, H., Sheth, F., Kumar, D., Gupta, A.K.: A computational intelligence framework integrating data augmentation and meta-heuristic opti- mization algorithms for enhanced hybrid nanofluid density prediction through machine and deep learning paradigms. IEEE Acces...

    2025

  22. [30]

    UCI Machine Learning Repository

    Whiteson, D.: HIGGS. UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C5V312 (2014)

    2014 doi

  23. [31]

    https://api.nasa.gov/

    Archive, N.E.: Kepler Exoplanet Dataset. https://api.nasa.gov/. Accessed: 2025- 05-10 (2025)

    2025

  24. [32]

    UCI Machine Learning Repository

    Bock, R.: MAGIC Gamma Telescope. UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C52C8B (2004)

    2004 doi

  25. [33]

    In: Chapelle, O., Chang, Y., Liu, T.-Y

    Chapelle, O., Chang, Y.: Yahoo! learning to rank challenge overview. In: Chapelle, O., Chang, Y., Liu, T.-Y. (eds.) Proceedings of the Learning to Rank Challenge. Proceedings of Machine Learning Research, vol. 14, pp. 1–24. Haifa, Israel (2011)

    2011

  26. [34]

    Journal of the American Statistical Association 106(494), 594–607 (2011) 33

    Cai, T., Liu, W., and, X.L.: A constrained ℓ1 minimization approach to sparse precision matrix estimation. Journal of the American Statistical Association 106(494), 594–607 (2011) 33

    2011

  27. [35]

    The Annals of Statistics 35(6), 2313–2351 (2007) 34

    CANDES, E., TAO, T.: The dantzig selector: Statistical estimation when p is much larger than n. The Annals of Statistics 35(6), 2313–2351 (2007) 34

    2007

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.