Pith. sign in

REVIEW 3 major objections 6 minor 31 references

Explaining Concept Shift with Interpretable Feature Attribution

T0 review · 3 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 SGShift attributes concept shift to a sparse set of features by fitting a penalized GAM update term, recovering most true shifted features in tests and ranking respiratory failure as the top COVID-19 shift on real data.

desk verdict A useful sparse concept-shift diagnostic, with an overreaching abstract and a theory that outruns the experiments. read the letter →

arxiv 2505.20634 v2 pith:LAAUEFVP submitted 2025-05-27 cs.LG stat.ML

Ruiqi Lyu , Alistair Turcan , Bryan Wilder This is my paper
classification cs.LGstat.ML
keywords conceptshiftdistributionfeatureattributiongeneralizedadditivemodelsknockoffsselectionfalsediscoverycontrolinterpretablemachinelearning
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

Concept shift happens when the relationship between features and labels changes between domains, and it is a major cause of model degradation in practice. SGShift treats explaining that shift as a feature-selection problem: it fits a sparse generalized additive update term on top of the source model and selects the few features whose coefficients are nonzero. The paper argues this recovers the genuinely shifted features far better than existing baselines, often with AUC above 0.9 and recall above 80-90% in semi-synthetic experiments, while needing only modest amounts of target-domain data. On real COVID-19 hospitalization data, the method ranks respiratory failure as the top shifted feature with a negative sign, matching the documented reduction in severity during Omicron. The result matters because it turns the vague fact that 'the data changed' into a specific, inspectable list of features to fix or investigate.

What carries the argument

The load-bearing object is the sparse GAM update term: an additive correction $\phi(X)^\top\delta$ placed on top of the frozen source model $\hat f$, so that all shift information lives in $\delta$. Because the basis functions are fixed, the problem becomes a sparse linear regression, which lets the authors bring in standard tools: $\ell_1$ regularization for selection, a restricted-strong-convexity analysis for the convergence guarantee, Model-X knockoffs for false-discovery control, and stability selection over repeated knockoff draws to derandomize the selection. The absorption variant adds a difference-in-difference structure in which coefficients $\omega$ appear in both domains and $\delta$ only in the target, so $\delta$ is not contaminated by poor fit of the source model.

What would settle it

Take a dataset where the true shift is constructed to be dense, with every coefficient in $\delta$ nonzero, or to involve only pairwise interactions, then run SGShift and its knockoff variants and check whether the reported sparse support still has high recall of the true shifted features; a method that returns a confident small list in that setting would refute the core attribution claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that concept shift can be attributed to a sparse set of features by modeling the target conditional mean as $g(E_T(y|X))=\hat f(X)+\phi(X)^\top\delta$, where $\hat f$ is the source-trained model and $\delta$ is a sparse coefficient vector over known basis functions. The shifted features are the support $A=\{j:\delta_j\neq 0\}$. SGShift estimates $\delta$ by minimizing the GAM loss on target data with an $\ell_1$ penalty, and the paper proves a lasso-type convergence bound under restricted strong convexity. Two extensions are introduced: an absorption term that fits shared coefficients on source and target data to absorb misspecification in $\hat f$, and a knockoff-based selection procedure that controls false discoveries, derandomized by stability selection over multiple knockoff draws. In semi-synthetic experiments on three healthcare datasets the knockoff variants achieve AUC over 0.9 in most settings and recall often above 80-90%, and on the real COVID-19 data the highest-ranked feature is respiratory failure with a negative sign.

Load-bearing premise

The central assumption is that the true concept shift is exactly a sparse linear combination of the chosen basis functions added to the source model; if the real shift is dense, nonlinear, or involves interactions not in the basis, the selected features can be incomplete or spurious.

Editorial extensions

If this is right

  • SGShift gives model developers a diagnosis step: when a deployed model loses accuracy on a new domain, the selected features are the ones to inspect, retrain, or fix first.
  • The knockoff variants carry a formal false-discovery guarantee, so users can report a feature list with a stated error rate rather than an uncalibrated ranking.
  • The method works with a small target sample and across model classes, including trees, linear models, SVMs, and boosting, and it tolerates misspecification of the source model through the absorption term.
  • On real healthcare data, updating only the features SGShift selects recovers a large share of the loss from training on the target alone, consistent with the claim that true concept shifts are often sparse.
  • The COVID-19 result shows that the attribution can be scientifically meaningful, singling out a feature whose sign agrees with known biology of variant severity.

Reading between the lines

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

  • The sparse-additive framing suggests a direct testable extension: apply SGShift to shifts that are deliberately dense or interaction-only; if it still selects a small set with high recall, the attribution is suspect, and if it fails, the method's scope is exactly the declared model class.
  • The difference-in-difference absorption idea could be reused in other attribution tasks, for example to separate confounding covariate drift from genuine label-relationship change when building shift reports.
  • Because SGShift outputs signed coefficients, it could be used for hypothesis generation about mechanism, such as a negative coefficient meaning a feature became protective, though causal claims would require additional identification assumptions.
  • The loss-recovery curves suggest a natural next step: using the identified shifted features to drive targeted model updates, since updating only the selected features can match or exceed target-only retraining.
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

3 major / 6 minor

Summary. The paper proposes SGShift, a method that attributes concept shift in tabular data to a sparse set of features by fitting a GAM update term ϕ(X)^T δ to the target labels, with three extensions: an absorption term for model misspecification (SGShift-A), knockoff-based false-discovery control (SGShift-K), and their combination (SGShift-KA). The central claim is that SGShift identifies shifted features much more accurately than baseline methods in semi-synthetic simulations, where it often reaches AUC above 0.9 and recall above 0.80, and that on real COVID-19, diabetes, and SUPPORT2 data the selected features align with clinical knowledge. The paper also states a convergence guarantee for the ℓ1-penalized estimator (Theorem 4.1) and FDR/PFER control for stability-selected knockoffs (Lemma 4.2).

Significance. If the claims hold, SGShift would be a practically useful diagnostic tool: it produces interpretable, sparse explanations for conditional distribution shift, works with black-box source models, and appears to require relatively few target samples. The semi-synthetic protocol preserves real feature correlations, the real-data finding on respiratory failure and Omicron severity is an external check against the medical literature, and the paper ships per-model-configuration results and code. The main limitations are that the empirical evaluation draws shifts only from the method's own sparse additive generative family, and that the formal proofs for Theorem 4.1 and Lemma 4.2 contain gaps. The contribution is therefore promising but currently overclaimed relative to what is established.

major comments (3)
  1. [Section 3.2, Section 5, Table 2] The evaluation only tests concept shifts generated inside the model assumption g(E_T(y|X)) = f(X) + ϕ(X)^T δ with δ sparse, so the central claim that SGShift identifies shifted features under general concept shift is not yet supported. The semi-synthetic design perturbs g(E[y|X]) through additive transformations based on selected features (Section 5), which is exactly the estimator's function class; the basis ϕ is never specified in the main text and no experiment varies it. Dense shifts, shifts involving interactions absent from ϕ, or shifts whose support is not identifiable under feature correlation could produce incomplete or spurious selections, and the discussion concedes that dense shifts 'may exist naturally'. Figure 2's elbow is computed from features already selected by SGShift and cannot by itself certify that natural shifts are sparse in the chosen basis. The authors should add experiments with dense δ, with interaction or nonlinear basis misspecification, and with different explicit choices of ϕ, or explicitly restrict the claim to shifts representable in the assumed sparse additive family.
  2. [Theorem 4.1 (Section 4.1, Appendix A.2)] The proof of the convergence guarantee is not valid as written. Condition (3) states n_T λ = λ' ≍ sqrt(log K / n_T), which gives λ ≍ sqrt(log K) / n_T^{3/2}, but the proof uses λ ≍ sqrt(log K / n_T); these scalings are inconsistent. More importantly, the proof simply assumes the cone condition ∥d_{A^c}∥_1 ≤ ∥d_A∥_1 in 'standard Lasso analysis', but this condition is not derived from the stated assumptions; it typically requires a subgradient bound with c_3 < 1 or an additional restricted-eigenvalue argument. Even granting the RSC and subgradient assumptions, the displayed derivation does not establish the claimed error bound. The theorem should be repaired by stating the correct scaling, proving the cone condition, and verifying or explicitly assuming any additional constants that are needed.
  3. [Lemma 4.2 (Section 4.3, Appendix A.4)] The FDR-control portion of Lemma 4.2 is not proven by the displayed argument. The chain starts with FDR(π) ≤ E[Σ_b V^b / (Bπ)], then claims an equality involving E[R^b E[V^b/(R^b∨1)|R^b]], and after applying the geometric thinning inequality concludes FDR ≤ q/(1-(1-π)^B). The last substitution does not follow: the geometric thinning inequality gives Σ_b E[R^b] ≥ E[|A(π)|]/(1-(1-π)^B), which introduces the extra factor E[|A(π)|] and leaves a bound that is not the stated one. The proof also requires a conditional version of the per-iteration FDR control that is not stated or derived. A correct proof, or a direct citation to the precise derandomized-knockoffs result in [RWC23] that covers the offset/update setting, is needed before the paper can claim rigorous FDR control.
minor comments (6)
  1. [Abstract] The opening sentence contains two grammatical errors: 'Regardless the amount' should be 'Regardless of the amount', and 'to have learned' should read 'having learned'.
  2. [Section 4.3 / Appendix A.3] There is a typo in the knockoff description: 'leaves the koint distribution' should be 'leaves the joint distribution'.
  3. [Section 4.3] The knockoff threshold τ is introduced but it is not specified how τ is chosen in practice to achieve a given per-iteration FDR or PFER level; a concrete selection rule should be stated.
  4. [Section 6, Table 2] The caption says results are aggregated across matched and mismatched configurations and that the highest performing method is bold, but the table presents separate Matched and Mismatched rows with bold entries in each; please clarify the aggregation and the bold-face rule.
  5. [Section 3.2 and throughout] The basis functions ϕ(X) are never defined; the manuscript should state what these functions are (e.g., spline bases for each feature, linear terms, or another explicit class) since they determine what 'shifted features' means in the method.
  6. [Figure 2] The caption states 'Each dot represents a penalty between 0 and 1', but the direction in which the penalty changes (increasing vs. decreasing λ) and how the number of selected features is determined should be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is self-contained and real-data validation is external to the method.

full rationale

SGShift's derivation chain is self-contained. Section 3.2 posits a sparse additive shift model g(E_T(y|X)) = f(X) + phi(X)^T delta with delta sparse; the estimator in Section 4.1 minimizes a penalized likelihood for delta. The convergence guarantee (Theorem 4.1) and FDR control (Lemma 4.2) rest on standard external results (RSC, model-X knockoffs, derandomized knockoffs [CFJL18, RWC23, RV10, RWY10]) rather than on the method's own conclusions. The semi-synthetic experiments construct ground-truth shifts by the same additive form used in the estimator; this is a standard simulation design and does not make the recovered support equal to the input by construction, because the estimation problem remains nontrivial and is evaluated against known labels. Real-data findings are checked against independent medical literature (e.g., Omicron severity), and the paper explicitly concedes that dense concept shifts may exist naturally. The only self-referential element is Figure 2, which uses SGShift's own selected features to argue that real shifts are sparse; this is an empirical demonstration that the paper does not promote to a formal derivation, and it is accompanied by an explicit limitation. No load-bearing self-citations or imported uniqueness theorems are present.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the sparse additive basis assumption for the shift, standard GLM and GAM assumptions, and, for the theoretical guarantees, sub-Gaussian designs, RSC, and a subgradient bound that are stated but not verified. Model-X knockoffs add a further assumption that the feature distribution is correctly modeled. The free parameters are the tuning hyperparameters (penalties, knockoff threshold, stability threshold) whose operating points are selected by sweeping rather than a principled rule.

free parameters (4)
  • l1 penalty lambda for shift coefficients delta = swept over [0,1] in experiments, no principled selection rule reported
    Controls the sparsity and size of the estimated shift; the paper evaluates along the penalty path rather than selecting a single operating point.
  • absorption penalties lambda_omega and lambda_delta = not stated in main text; only constraint lambda_omega < lambda_delta given
    Hierarchical penalty separates model misspecification (omega) from target-specific shift (delta); values are user-chosen.
  • knockoff threshold tau = not stated
    Determines per-iteration FDR/PFER control in SGShift-K; the paper does not specify how tau is set.
  • stability threshold pi = user-determined parameter
    Defines the selection frequency cutoff in derandomized knockoffs; no default is given.
assumptions (5)
  • domain assumption The concept shift is representable as g(E_T(y|X)) = f(X) + phi(X)^T delta with delta sparse
    Stated in Section 3.2 as the core modeling assumption; the paper's experiments generate shifts this way, so violations are untested.
  • domain assumption The response follows an exponential family with canonical link g(.)
    Used to define the negative log-likelihood loss in Section 3.2; standard GLM assumption.
  • standard math The design vectors phi_i are i.i.d. sub-Gaussian with positive definite covariance
    Used in Lemma .1 to establish RSC and in Theorem 4.1; standard in high-dimensional statistics but unverified for real tabular features.
  • domain assumption Restricted Strong Convexity and the subgradient bound ||grad L(delta*)||_inf <= c3 lambda hold
    Assumptions 1 and 2 of Theorem 4.1 are stated without verification; the subgradient bound also requires the source model f_hat to be sufficiently close to the true f, which is not guaranteed.
  • domain assumption The Model-X knockoff distribution correctly models the joint distribution of features
    Needed for FDR control in Section 4.3; the paper does not describe how knockoffs are constructed for mixed categorical and continuous data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explaining Concept Shift with Interpretable Feature Attribution." pith.science (2026). https://pith.science/paper/LAAUEFVP

@misc{pith2026250520634,
  author       = {Pith},
  title        = {Pith review of: Explaining Concept Shift with Interpretable Feature Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LAAUEFVP}},
  note         = {Machine review of arXiv:2505.20634}
}
read the original abstract

Concept shift occurs when the distribution of labels conditioned on the features changes between domains, which can make even a well-tuned ML model miscalibrated on a new domain. Identifying these shifted features provides unique insight into how feature-label relationships differ between domains, considering the difference may be across a scientifically relevant dimension, such as time, disease status, population, etc. In this paper, we propose SGShift, a method for attributing performance degradation under concept shift in tabular data to a sparse set of shifted features. We frame concept shift as a feature selection task to learn the features that can explain performance differences between models in the source and target domain. This framework enables SGShift to adapt powerful statistical tools such as generalized additive models, knockoffs, and absorption towards identifying these shifted features. We conduct extensive experiments in synthetic and real data across various ML models and find SGShift can identify shifted features much more accurately than baseline methods, requires few samples in the shifted domain, and is robust to complex cases of concept shift.

Figures

Figures reproduced from arXiv: 2505.20634 by the authors.

Figure 1
Figure 1. Evaluation of recall at FPR 10% in semi-synthetic simulations for matched and mismatched [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Performance increase as we update more features in the source model identified by SGShift. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Real data results showing the ordering of selected features for each model as the penalty [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Individual model configuration results for diabetes readmission semi-synthetic simulations, [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Individual model configuration results for COVID-19 semi-synthetic simulations, measured [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: Individual model configuration results for SUPPORT2 semi-synthetic simulations, measured [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: Performance increase as we update more features in the source model identified by SGShift. [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 29 canonical work pages

  1. [1]

    Mortality risk among patients hospitalized primarily for covid-19 during the omicron and delta variant pandemic periods — united states, april 2020–june 2022

    Solomon Adjei, Kai Hong, Nicole M Molinari, et al. Mortality risk among patients hospitalized primarily for covid-19 during the omicron and delta variant pandemic periods — united states, april 2020–june 2022. MMWR Morbidity and Mortality Weekly Report , 71:1182--1189, 2022

    work page 2020

  2. [2]

    A controlled trial to improve care for seriously iii hospitalized patients: The study to understand prognoses and preferences for outcomes and risks of treatments (support)

    Alfred F Connors, Neal V Dawson, Norman A Desbiens, William J Fulkerson, Lee Goldman, William A Knaus, Joanne Lynn, Robert K Oye, Marilyn Bergner, Anne Damiano, et al. A controlled trial to improve care for seriously iii hospitalized patients: The study to understand prognoses and preferences for outcomes and risks of treatments (support). Jama , 274(20):...

    work page 1995

  3. [3]

    Panning for gold:‘model-x’knockoffs for high dimensional controlled variable selection

    Emmanuel Candes, Yingying Fan, Lucas Janson, and Jinchi Lv. 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

    work page 2018

  4. [4]

    Diagnosing model performance under distribution shift

    Tiffany Tianhui Cai, Hongseok Namkoong, and Steve Yadlowsky. Diagnosing model performance under distribution shift. arXiv preprint arXiv:2303.02011 , 2023

    arXiv 2023

  5. [5]

    Estimating and explaining model performance when both covariates and labels shift

    Lingjiao Chen, Matei Zaharia, and James Y Zou. Estimating and explaining model performance when both covariates and labels shift. Advances in Neural Information Processing Systems , 35:11467--11479, 2022

    work page 2022

  6. [6]

    A survey of methods for explaining black box models

    Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, Franco Turini, Fosca Giannotti, and Dino Pedreschi. A survey of methods for explaining black box models. ACM computing surveys (CSUR) , 51(5):1--42, 2018

    work page 2018

  7. [7]

    Benchmarking distribution shift in tabular data with tableshift

    Josh Gardner, Zoran Popovic, and Ludwig Schmidt. Benchmarking distribution shift in tabular data with tableshift. Advances in Neural Information Processing Systems , 2023

    work page 2023

  8. [8]

    Generalized additive models

    Trevor J Hastie. Generalized additive models. Statistical models in S , pages 249--307, 2017

    work page 2017

Show all 31 references
  1. [9]

    Omicron subvariant ba.5 efficiently infects lung cells

    Markus Hoffmann, Lanying YR Wong, Prerna Arora, et al. Omicron subvariant ba.5 efficiently infects lung cells. Nature Communications , 14:3500, 2023

    2023

  2. [10]

    Sequential covariate shift detection using classifier two-sample tests

    Sooyong Jang, Sangdon Park, Insup Lee, and Osbert Bastani. Sequential covariate shift detection using classifier two-sample tests. In International conference on machine learning , pages 9845--9880. PMLR, 2022

    2022

  3. [11]

    Feature shift detection: Localizing which features have shifted via conditional distribution tests

    Sean Kulinski, Saurabh Bagchi, and David I Inouye. Feature shift detection: Localizing which features have shifted via conditional distribution tests. Advances in neural information processing systems , 33:19523--19533, 2020

    2020

  4. [12]

    Towards explaining distribution shifts

    Sean Kulinski and David I Inouye. Towards explaining distribution shifts. In International Conference on Machine Learning , pages 17931--17952. PMLR, 2023

    2023

  5. [13]

    Targeting underrepresented populations in precision medicine: A federated transfer learning approach

    Sai Li, Tianxi Cai, and Rui Duan. Targeting underrepresented populations in precision medicine: A federated transfer learning approach. The annals of applied statistics , 17(4):2970, 2023

    2023

  6. [14]

    Transfer learning for high-dimensional linear regression: Prediction, estimation and minimax optimality

    Sai Li, T Tony Cai, and Hongzhe Li. Transfer learning for high-dimensional linear regression: Prediction, estimation and minimax optimality. Journal of the Royal Statistical Society Series B: Statistical Methodology , 84(1):149--173, 2022

    2022

  7. [15]

    Clinical outcomes associated with sars-cov-2 omicron (b.1.1.529) variant and ba.1/ba.1.1 or ba.2 subvariant infection in southern california

    Joseph A Lewnard, Vincent X Hong, Manish M Patel, et al. Clinical outcomes associated with sars-cov-2 omicron (b.1.1.529) variant and ba.1/ba.1.1 or ba.2 subvariant infection in southern california. Nature Medicine , 28:1933--1943, 2022

    1933

  8. [16]

    On the need for a language describing distribution shifts: Illustrations on tabular datasets

    Jiashuo Liu, Tianyu Wang, Peng Cui, and Hongseok Namkoong. On the need for a language describing distribution shifts: Illustrations on tabular datasets. Advances in Neural Information Processing Systems , 36:51371--51408, 2023

    2023

  9. [17]

    Rethinking distribution shifts: Empirical analysis and inductive modeling for tabular data, 2024

    Jiashuo Liu, Tianyu Wang, Peng Cui, and Hongseok Namkoong. Rethinking distribution shifts: Empirical analysis and inductive modeling for tabular data, 2024

    2024

  10. [18]

    Detecting and correcting for label shift with black box predictors

    Zachary Lipton, Yu-Xiang Wang, and Alexander Smola. Detecting and correcting for label shift with black box predictors. In International conference on machine learning , pages 3122--3130. PMLR, 2018

    2018

  11. [19]

    Stability selection

    Nicolai Meinshausen and Peter B \"u hlmann. Stability selection. Journal of the Royal Statistical Society Series B: Statistical Methodology , 72(4):417--473, 2010

    2010

  12. [20]

    Explanation shift: How did the distribution shift impact the model? arXiv preprint arXiv:2303.08081 , 2023

    Carlos Mougan, Klaus Broelemann, David Masip, Gjergji Kasneci, Thanassis Thiropanis, and Steffen Staab. Explanation shift: How did the distribution shift impact the model? arXiv preprint arXiv:2303.08081 , 2023

    2023 arXiv

  13. [21]

    Interpretable generalized additive models for datasets with missing values

    Hayden McTavish, Jon Donnelly, Margo Seltzer, and Cynthia Rudin. Interpretable generalized additive models for datasets with missing values. Advances in Neural Information Processing Systems , 37:11904--11945, 2025

    2025

  14. [22]

    all of us

    All of Us Research Program Investigators. The “all of us” research program. New England Journal of Medicine , 381(7):668--676, 2019

    2019

  15. [23]

    Non-asymptotic theory of random matrices: extreme singular values

    Mark Rudelson and Roman Vershynin. Non-asymptotic theory of random matrices: extreme singular values. In Proceedings of the International Congress of Mathematicians 2010 (ICM 2010) (In 4 Volumes) Vol. I: Plenary Lectures and Ceremonies Vols. II--IV: Invited Lectures , pages 15...

    2010

  16. [24]

    Derandomizing knockoffs

    Zhimei Ren, Yuting Wei, and Emmanuel Cand \`e s. Derandomizing knockoffs. Journal of the American Statistical Association , 118(542):948--958, 2023

    2023

  17. [25]

    Restricted eigenvalue properties for correlated gaussian designs

    Garvesh Raskutti, Martin J Wainwright, and Bin Yu. Restricted eigenvalue properties for correlated gaussian designs. The Journal of Machine Learning Research , 11:2241--2259, 2010

    2010

  18. [26]

    Impact of hba1c measurement on hospital readmission rates: analysis of 70,000 clinical database patient records

    Beata Strack, Jonathan P DeShazo, Chris Gennings, Juan L Olmo, Sebastian Ventura, Krzysztof J Cios, and John N Clore. Impact of hba1c measurement on hospital readmission rates: analysis of 70,000 clinical database patient records. BioMed research international , 2014(1):781670, 2014

    2014

  19. [27]

    Covariate shift adaptation by importance weighted cross validation

    Masashi Sugiyama, Matthias Krauledat, and Klaus-Robert M \"u ller. Covariate shift adaptation by importance weighted cross validation. Journal of Machine Learning Research , 8(5), 2007

    2007

  20. [28]

    High-dimensional generalized linear models and the lasso

    Sara A Van de Geer. High-dimensional generalized linear models and the lasso. 2008

    2008

  21. [29]

    High-dimensional probability: An introduction with applications in data science , volume 47

    Roman Vershynin. High-dimensional probability: An introduction with applications in data science , volume 47. Cambridge university press, 2018

    2018

  22. [30]

    Wickenhagen, M

    A. Wickenhagen, M. Flagg, J.R. Port, et al. Evolution of omicron lineage towards increased fitness in the upper respiratory tract in the absence of severe lung pathology. Nature Communications , 16:594, 2025

    2025

  23. [31]

    why did the model fail?

    Haoran Zhang, Harvineet Singh, Marzyeh Ghassemi, and Shalmali Joshi. " why did the model fail?": Attributing model performance changes to distribution shifts. 2023

    2023

Pith tools

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