Pith. sign in

REVIEW 2 major objections 4 minor 64 references

Optimal Empirical Risk Minimization under Temporal Distribution Shifts

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Under a random-shift model of drifting data, the best weights for past observations minimize a two-term trade-off: residual distributional change plus sampling noise, scaled by problem difficulty.

desk verdict The core idea is solid and the special-case derivations are genuinely new, but Theorem 1 as stated has a normalization error that the proof does not support, and the abstract overclaims the stock volatility results. read the letter →

arxiv 2507.13287 v1 pith:3DIZEA4U submitted 2025-07-17 stat.ME cs.LG

Yujin Jeong , Ramesh Johari , Dominik Rothenhäusler , Emily Fox This is my paper · ORCID
classification stat.MEcs.LG MSC 62F1262M10
keywords temporaldistributionshiftweightedempiricalriskminimizationrandommodelout-of-distributiongeneralizationtimeseriesweightingexponentialdown-weightingrealizedvolatilityforecastingfine-tuning
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 claims that under a random-shift model of temporal drift, there is a provably optimal way to weight past data in empirical risk minimization (ERM): choose weights that minimize residual distributional variation plus sampling variation, scaled by problem difficulty. If the claim is right, three weighting strategies in everyday use — pooling all past data, exponentially down-weighting older data, and using only the most recent data — are not heuristics but exact limiting cases of one formula. The paper derives the formula asymptotically, provides parametric and nonparametric estimators for the weights that need no assumptions on the data distributions themselves, and reports that the method improves out-of-sample performance as a fine-tuning step on the Wild-Time Yearbook benchmark, and on stock volatility and NYC taxi ride-duration forecasting.

What carries the argument

The load-bearing object is the excess-risk decomposition of Theorem 1, which separates the out-of-distribution error of any fixed weighting $\beta$ into residual distributional variation $\mathbb{E}[(W_{T+1} - \sum_k \beta_k W_{T+1-k})^2]$, sampling variation $\sum_k \beta_k^2 r_{T+1-k}$, and a shared difficulty multiplier $\mu = \mathrm{Tr}(\mathbb{E}_0[\partial_\theta^2 L(\theta_0,D)]^{-1} \mathrm{Var}_{P_0}(\partial_\theta L(\theta_0,D)))$. Equation (1) minimizes the first two terms, and the weakly stationary time-series model of the weights (AR, MA, ARMA) converts that minimization into closed-form parametric weights. Theorem 2 then supplies a distributional central limit theorem showing that the variance-inflation factor $\tilde{\delta}^2(\beta) = \delta^2(\beta) + r\beta^{\top}\beta$ scales the limiting covariance of any test function, which is what the nonparametric estimator minimizes to recover optimal weights from data.

What would settle it

Construct a data-generating process whose true likelihood-ratio shifts are concentrated in a single bin (or a few correlated coordinates) instead of spread independently across all $m$ bins, and compare the one-step-ahead excess risk of the Eq. (1) weights against simple pooling and most-recent weighting in the $r \to 0$ regime as $m$ grows; if RIDER's weights do not match or beat those baselines, the Theorem 1 decomposition has not identified the right minimizer for localized shifts.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 1: under the random distribution shift model, the one-step-ahead excess risk of weighted ERM is asymptotically proportional to $\mathbb{E}[(W_{T+1} - \sum_{k=1}^K \beta_k W_{T+1-k})^2] + \sum_{k=1}^K \beta_k^2 r_{T+1-k}$, multiplied by a problem-difficulty constant $\mu$. The first term is residual distributional variation and the second is sampling variation, so the asymptotically optimal weights solve $\beta^* = \arg\min_{\beta \geq 0,\, \beta^{\top}\mathbf{1}=1} \delta^2(\beta) + \sum_k \beta_k^2 r_{T+1-k}$, trading off two sources of error that earlier formulations did not balance. The model posits that the likelihood ratio $dP_t/dP_0$ is piecewise constant on $m$ bins with i.i.d. positive random weights of mean 1, making residual shifts dense and idiosyncratic; modeling the weight process as weakly stationary turns weight selection into a time-series problem. From this, the paper derives that pooling ($r \to \infty$), using only the most recent data (AR(1) weights with $r \to 0$), and exponential down-weighting (ARMA(1,1) weights with $r \to 0$) are exact limits of the optimal weighting.

Load-bearing premise

The load-bearing premise is the random distribution shift model of Section 2.2: residual shifts must act as dense, independent perturbations — i.i.d. positive random weights across many bins with finite variance — whose time evolution is weakly stationary; the paper itself calls the i.i.d.-across-bins assumption strong, and if real shifts are localized in a few components or strongly dependent across them, the optimal-weight formula and the derived weight shapes need not hold.

Editorial extensions

If this is right

  • If the random-shift model holds, RIDER's weights are asymptotically optimal for one-step-ahead prediction under weighted ERM, with no parametric assumption on the data-generating distributions themselves.
  • Pooling all data is optimal exactly when sampling noise dominates distributional movement ($r \to \infty$); exponential down-weighting and most-recent-only weighting emerge as limits of ARMA(1,1) and AR(1) weight processes when distributional movement dominates ($r \to 0$).
  • Because the weights are estimated from pairwise similarity of datasets via test functions, RIDER can be layered on top of any ERM-based learner; the paper reports consistent accuracy gains when used as a fine-tuning step across the Wild-Time Yearbook benchmark.
  • The variance-inflation identity implies the estimated weights transfer across loss functions and models, since changing the loss only changes the scalar $\mu$, not the optimal weighting.

Reading between the lines

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

  • A direct test of the model would estimate the weight autocovariance from data with labeled time splits and compare the predicted optimal weights with cross-validated weights; systematic mismatch would indicate residual shifts that are not dense and weakly stationary.
  • The loss-agnostic form of the variance-inflation factor suggests the same weighting procedure could apply to non-ERM objectives such as quantile or likelihood-based forecasting, which the paper does not explore.
  • Relaxing weak stationarity to local stationarity of the weight process would yield time-varying optimal weights, interpolating between RIDER and purely recency-based schemes; this generalization is left implicit in the paper.
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

2 major / 4 minor

Summary. The paper proposes RIDER, a weighted empirical risk minimization method for one-step-ahead prediction under temporal distribution shifts. It models residual shifts as random, dense perturbations of a parent distribution via i.i.d. bin weights with weak temporal stationarity. The main theoretical result, Theorem 1, gives an asymptotic decomposition of the excess risk into residual distributional variation and sampling variation, leading to an optimal-weight optimization problem (1). The paper derives pooling, most-recent, and exponential weights as limiting special cases under AR/ARMA models, and proposes parametric and nonparametric weight estimators with consistency results. Experiments apply RIDER as a fine-tuning layer on Wild-Time/Yearbook and as a standalone method for stock volatility and NYC taxi ride duration.

Significance. If Theorem 1 is repaired to the normalization actually proven, the paper offers a useful theoretical bridge between weighted ERM and standard time-series downweighting heuristics; the explicit trade-off between sampling and distributional variability is a genuine contribution. The nonparametric CLT and consistency result provide a principled route to weight estimation, and the Yearbook fine-tuning gains are consistent across many baselines. The taxi experiment shows practical gains from outcome-aware test functions. The main caveats are that the stock volatility experiment only matches exponential weighting rather than outperforming it, and the whole framework rests on the strong i.i.d.-across-bins random-shift assumption, which the authors themselves acknowledge.

major comments (2)
  1. [Section 2.3, Theorem 1 and Appendix D.3] The stated normalization in Theorem 1 is inconsistent with its proof. The proof establishes in Eq. (8) that sqrt(m)(hatθ^β_t - θ_t) converges to a Gaussian limit, and then in Eq. (9) multiplies the excess risk by m, not by sqrt(m); hence m times the excess risk converges to a nondegenerate quadratic form with the stated mean. If sqrt(m) times the excess risk were the object, it would equal (m times the excess risk)/sqrt(m), which converges to zero in probability. The theorem statement should therefore read m(N) times the excess risk, not sqrt(m(N)). Because Eq. (1) and the subsequent optimal-weight special cases depend on the mean of this limit, the corrected statement preserves the paper's main conclusions, but the theorem as typeset is false and must be fixed.
  2. [Appendix D.1, proof of Lemma 3] The Berry-Esseen bound in Lemma 3 is written with a factor sqrt(N): ψ0 = [bracket] * sqrt(N). Since the bracketed ratio converges in probability to a finite positive constant, this bound would diverge, not vanish, as N → ∞. The displayed expression should presumably be 1/sqrt(N) to yield the claimed o(1). As written the proof of Lemma 3 is invalid, and since Lemma 3 supports Theorem 2 and hence Theorem 1, this normalization needs to be corrected.
minor comments (4)
  1. [Abstract and Section 4.2] The abstract and Section 4.2 overstate the stock volatility result: Figure 2 shows RIDER is statistically indistinguishable from exponential weighting (the rightmost boxplot has p = 0.57), so the claim that RIDER 'outperforms standard weighting strategies' in that task should be qualified.
  2. [Appendix D.3, proof of Theorem 1] The proof of Theorem 1 invokes Lemma 2 for functions such as ∂²θ L(θ0,D) and φ(D), which are square-integrable by Assumption 1 but not necessarily bounded; Lemma 2 is stated only for bounded functions. The L2 extension proved in Theorem 2 should be cited explicitly at that step.
  3. [Section 4.2] The stock volatility procedure adds constraints β1 ≤ B and monotonicity, with B chosen by cross-validation against exponential weights, that do not appear in the theoretical problem (1); this makes the empirical comparison with exponential weighting less direct and should be discussed.
  4. [Section 3 and Section 4.2] The empirical 'recovery' of exponential weighting in Section 4.2 is partly built into the parametric family (2), which already contains an exponential component; the non-circular support comes from Case 3 in Section 2.4, and the text should make this distinction explicit.

Circularity Check

2 steps flagged · score 4.0 of 10

Central derivation is not circular — optimal weights genuinely minimize an explicit asymptotic-risk criterion — but the empirical 'recovery of exponential weighting' is partly built into the fitted parametric family (2) and the stock-experiment constraints; separately, Theorem 1 as typeset contradicts its own m-normalized proof, a correctness flaw, not circularity.

  1. fitted input called prediction [Section 3, Eq. (2); Section 4.2 (stock experiment constraints and the 'recover exponential weighting' claim)]
    "β∗k = α1 · 1/K + α2 · I(k = 1) + α3 · θ^{k−1}/Σ_{k=1}^K θ^{k−1} ... The upper bound B is chosen based on reference exponential weights in (3), using various half-life values. The half-life value is chosen via cross-validation. ... In fact, we recover exponential weighting both theoretically and empirically."

    Pure exponential weighting is the special case of the parametric family (2) with α1 = α2 = 0, α3 = 1, and (α, θ) are fitted by cross-validation in the experiments; in the stock experiment the weights are additionally constrained monotone non-increasing, with β1 capped via 'B chosen based on reference exponential weights in (3)' and a cross-validated half-life. The claim of empirically recovering exponential weighting is therefore partly a consequence of the estimator's feasible set and the fitted family rather than an independent finding — the exponential-like weights are statistically facilitated by construction. This is a mild fitted-input-called-prediction.

  2. other [Theorem 1 (Section 2.3) vs. its proof (Appendix D.3, Eqs. (8)–(9)); Lemma 3 proof (Berry-Esseen factor)]
    "√m(N ) (E_{T+1}[L(θ̂β_{T+1}, D)] − E_{T+1}[L(θ_{T+1}, D)]) →d E for some random variable E, whose mean E[E] can be decomposed into (E[(W^{T+1} − Σ_{k=1}^K β_kW^{T+1−k})^2] + Σ_{k=1}^K β^2_k r_{T+1−k}) · µ"

    Flagged per the reviewing rule: an in-scope proof inconsistency, not a circularity. Eq. (8) gives √m(θ̂β_t − θ_t) = √m(Σ_k β_k Ê_k[φ] − E_t[φ]) + o_P(1), nondegenerate by Theorem 2. Eq. (9) writes the 'rescaled excess risk' as m·(E_t[L(θ̂β_t)] − E_t[L(θ_t)]) = m·(θ̂β_t − θ_t)⊺E_0[∂²_θL](θ̂β_t − θ_t) + o_P(1). Hence m·(excess risk) converges to a quadratic form with the mean δ̃²·µ claimed in Theorem 1, whereas √m·(excess risk) = (m·excess risk)/√m → 0 in probability. The stated √m(N) normalization cannot converge to a nondegenerate E; as typeset the central theorem is false, and the proof supports an m(N)-normalized statement. Program (1) rests on the mean formula the m-normalized proof establishes.

full rationale

Verdict: the core derivation is not circular. The optimal-weight program (1) is defined by an explicit criterion — the asymptotic mean δ²(β) + Σ β_k² r_{T+1−k} obtained from Theorem 1 — not by any assumed weighting scheme. Pooling (r→∞), most-recent (AR(1), r→0), and exponential (ARMA(1,1), r→0) weighting are derived as genuine limits (the ARMA(1,1) case via the infinite-order best-linear-prediction representation), and Section 3's estimator minimizes an empirical version of the same δ̃² objective, which Proposition 1 shows is consistent. The foundation is partly self-citational: the random-shift model and Lemma 1 are attributed to Jeong and Rothenhäusler (2024), which shares authors with this paper, and every downstream theorem imports Lemma 1's distributional CLT. However, the model's assumptions are restated fully in Section 2.2, the lemma is an auxiliary parameter-free CLT whose stated assumptions do not include the target result, and the paper openly flags the i.i.d.-across-bins assumption as strong and lists limitations in Section 6; per the rules this is real evidence, not circularity. The genuine circularity-adjacent finding is partial and empirical: the demonstration that RIDER 'recovers exponential weighting' is facilitated by construction because the fitted parametric family (2) contains exponential weights as a special case and the stock experiment's constraints (monotonicity; B chosen from exponential reference weights (3)) pre-shape the solution toward exponential-like vectors. Two non-circularity flaws are flagged per the reviewing rule: Theorem 1 as typeset states a √m(N) limit while its own proof (Eqs. (8)–(9)) establishes an m(N)-normalized limit, so as written the central theorem is false (the correct statement follows from the proof); and the Berry-Esseen bound in Lemma 3's proof displays a √N factor that should be 1/√N. These affect correctness, not circularity. Score 4: the central optimality claim retains independent theoretical content and is checked against external benchmarks (Wild-Time, stock, NYC taxi), so 6+ would be disproportionate; equal-weight, most-recent, and exponential baselines are all outperformed or matched on held-out data.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests primarily on the random distribution shift model and weak stationarity of the weight process; the practical estimator adds fitted hyperparameters (alpha, theta, K, B) and an assumed parametric family. No new physical or formal entities are postulated.

free parameters (4)
  • alpha_1, alpha_2, alpha_3 (mixture weights in Eq. (2)) = selected via cross-validation from grid {0, 0.2, ..., 1} on Yearbook
    These determine the balance among uniform, most-recent, and exponential components of the estimated weights; no closed-form estimator is given.
  • theta (decay rate in Eq. (2)) = selected via cross-validation over theta = (1/2)^(1/h), h in {2,4,6,8}
    The exponential down-weighting rate is a free hyperparameter fitted to data, not derived from ARMA parameters in the experiments.
  • window size K = cross-validated over {10,20} for Yearbook; {26,39,52,65,78} for stock volatility
    The number of past datasets included in the weighted ERM is chosen by cross-validation in every experiment.
  • upper bound B in stock volatility experiment = chosen via cross-validation based on exponential reference weights with half-life H in {6,9,12}
    A monotonicity and boundedness constraint is imposed on the weights to regularize estimation; B is not derived from the theory and affects the recovered weight shape.
assumptions (6)
  • standard math Dudley representation: any random variable/vector can be written as a measurable function of a uniform random variable on [0,1]
    Used in Section 2.2 to justify constructing distributional perturbations on the unit interval.
  • domain assumption Random distribution shift model: dP_t/dP_0 is piecewise constant on m bins and the weights W_{t,1},...,W_{t,m} are i.i.d. positive with E=1 and finite variance
    This is the generative model underlying Theorem 1 and Theorem 2; the paper itself calls the i.i.d. assumption strong in Section 2.2.
  • domain assumption Weak stationarity of the random weight process (Cov(W_t, W_{t+h}) = rho(h)) with AR/ARMA special cases
    Assumed throughout Section 2.4; needed to make delta^2 independent of t and to derive the exponential and most-recent special cases.
  • domain assumption Assumption 1 (regularity of the loss, nonsingular second moment, integrability) and Assumption 2 (sampling and distributional uncertainty of the same order, r in (0, infinity))
    Stated in Section 2.3; the asymptotic decomposition and the same-order trade-off depend on these.
  • ad hoc to paper Parametric family Eq (2) for estimated weights: beta_k = alpha1/K + alpha2 I(k=1) + alpha3 theta^{k-1}/sum theta^{k-1}
    This three-component family is not derived for general ARMA(p,q) processes; it is a modeling choice used to enable cross-validation-based estimation.
  • domain assumption Test functions phi_1,...,phi_L are uncorrelated with unit variance under P0, with L to infinity or T to infinity and decaying covariance under mixing
    Proposition 1's consistency of the nonparametric weight estimator requires this; Remark 1 suggests a whitening transform, but experiments do not report verifying it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimal Empirical Risk Minimization under Temporal Distribution Shifts." pith.science (2026). https://pith.science/paper/3DIZEA4U

@misc{pith2026250713287,
  author       = {Pith},
  title        = {Pith review of: Optimal Empirical Risk Minimization under Temporal Distribution Shifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3DIZEA4U}},
  note         = {Machine review of arXiv:2507.13287}
}
read the original abstract

Temporal distribution shifts pose a key challenge for machine learning models trained and deployed in dynamically evolving environments. This paper introduces RIDER (RIsk minimization under Dynamically Evolving Regimes) which derives optimally-weighted empirical risk minimization procedures under temporal distribution shifts. Our approach is theoretically grounded in the random distribution shift model, where random shifts arise as a superposition of numerous unpredictable changes in the data-generating process. We show that common weighting schemes, such as pooling all data, exponentially weighting data, and using only the most recent data, emerge naturally as special cases in our framework. We demonstrate that RIDER consistently improves out-of-sample predictive performance when applied as a fine-tuning step on the Yearbook dataset, across a range of benchmark methods in Wild-Time. Moreover, we show that RIDER outperforms standard weighting strategies in two other real-world tasks: predicting stock market volatility and forecasting ride durations in NYC taxi data.

Figures

Figures reproduced from arXiv: 2507.13287 by the authors.

Figure 1
Figure 1. Diagram illustrating the data-generating process under the random temporal distribution shift model [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. [Left] Boxplots showing the percentage differences in test MSEs of our method compared to three [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. [Left] Boxplots showing the percentage differences in test MSEs of our method compared to three [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 55 canonical work pages

  1. [1]

    T. Adel, H. Zhao, and R. E. Turner. Continual learning with adaptive weights (claw). arXiv preprint arXiv:1911.09514, 2019

    work page Pith review arXiv 1911

  2. [2]

    Arjovsky, L

    M. Arjovsky, L. Bottou, I. Gulrajani, and D. Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019

    arXiv 1907

  3. [3]

    G. Bai, L. Chen, and L. Zhao. Temporal domain generalization with drift-aware dynamic neural network. In International Conference on Learning Representations (ICLR), 2022

    work page 2022

  4. [4]

    K. C. Bansak, E. Paulson, and D. Rothenh \"a usler. Learning under random distributional shifts. In International Conference on Artificial Intelligence and Statistics, pages 3943--3951. PMLR, 2024

    work page 2024

  5. [5]

    P. L. Bartlett. Learning with a slowly changing distribution. In Proceedings of the Fifth Annual Workshop on Computational Learning Theory (COLT), 1992

    work page 1992

  6. [6]

    R. D. Barve and P. M. Long. On the complexity of learning from drifting distributions. In Conference on Computational Learning Theory (COLT), 1996

    work page 1996

  7. [7]

    Ben-Tal, D

    A. Ben-Tal, D. den Hertog, A. De Waegenaere, B. Melenberg, and G. Rennen. Robust mean-square error estimation. Mathematical Programming, 147 0 (1): 0 277--313, 2013

    work page 2013

  8. [8]

    Bennett and J

    S. Bennett and J. Clarkson. Time series prediction under distribution shift using differentiable forgetting. arXiv preprint arXiv:2207.11486, 2022

    arXiv 2022

Show all 64 references
  1. [9]

    D. M. Blei and J. D. Lafferty. Dynamic topic models. In Proceedings of the 23rd International Conference on Machine Learning, page 113–120. Association for Computing Machinery, 2006

    2006

  2. [10]

    Caron, I

    M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33: 0 9912--9924, 2020

    2020

  3. [11]

    Chaudhry, M

    A. Chaudhry, M. Ranzato, M. Rohrbach, and M. Elhoseiny. Efficient lifelong learning with a-gem. In International Conference on Learning Representations, 2019

    2019

  4. [12]

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597--1607. PMLR, 2020

    2020

  5. [13]

    Christie, N

    G. Christie, N. Fendley, J. Wilson, and R. Mukherjee. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  6. [14]

    C. B. Clement, M. Bierbaum, K. P. O’Keeffe, and A. A. Alemi. On the use of arxiv as a dataset. https://arxiv.org/abs/1905.00075, 2019. arXiv preprint arXiv:1905.00075

    1905 arXiv

  7. [15]

    Delage and Y

    E. Delage and Y. Ye. Distributionally robust optimization under moment uncertainty with application to data-driven problems. Operations Research, 58 0 (3): 0 595--612, 2010

    2010

  8. [16]

    J. C. Duchi and H. Namkoong. Learning models with uniform performance via distributionally robust optimization. In The Annals of Statistics, volume 49, pages 1378--1406, 2021

    2021

  9. [17]

    R. M. Dudley. Real analysis and probability. CRC Press, 2018

    2018

  10. [18]

    Ganin and V

    Y. Ganin and V. Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning (ICML), pages 1180--1189, 2015

    2015

  11. [19]

    Ginosar, K

    S. Ginosar, K. Rakelly, S. Sachs, B. Yin, and A. A. Efros. A century of portraits: A visual historical record of american high school yearbooks. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1--7, 2015

    2015

  12. [20]

    Gretton, J

    A. Gretton, J. Smola, M. Huang, K. Schmittfull, K. Borgwardt, and B. Sch \"o lkopf. Covariate shift by kernel mean matching. Dataset shift in machine learning, 3: 0 131--160, 2009

    2009

  13. [21]

    Gulrajani and D

    I. Gulrajani and D. Lopez-Paz. In search of lost domain generalization. In ICLR, 2021

    2021

  14. [22]

    L. L. Guo, S. R. Pfohl, J. Fries, A. E. W. Johnson, J. Posada, C. Aftandilian, N. Shah, and L. Sung. Evaluation of domain generalization and adaptation on improving model robustness to temporal dataset shift in clinical medicine. Scientific Reports, 12 0 (1): 0 1--10, 2022

    2022

  15. [23]

    Hanneke and L

    S. Hanneke and L. Yang. Statistical learning under nonstationary mixing processes. In International Conference on Artificial Intelligence and Statistics (AISTATS), 2019

    2019

  16. [24]

    Heber, A

    G. Heber, A. Lunde, N. Shephard, and K. Sheppard. Oxford-man institute of quantitative finance. realized library, 2013. Version: 0.2

    2013

  17. [25]

    W. Hu, G. Niu, I. Sato, and M. Sugiyama. Does distributionally robust supervised learning give robust classifiers? In International Conference on Machine Learning, pages 2029--2037. PMLR, 2018

    2018

  18. [26]

    Huang, T

    K. Huang, T. Fu, W. Gao, Y. Zhao, Y. Roohani, J. Leskovec, C. W. Coley, C. Xiao, J. Sun, and M. Zitnik. Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development. arXiv preprint arXiv:2102.09548, 2021

    2021 arXiv

  19. [27]

    Izmailov, D

    P. Izmailov, D. Podoprikhin, T. Garipov, D. Vetrov, and A. G. Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018

    2018 arXiv

  20. [28]

    Jeong and D

    Y. Jeong and D. Rothenhäusler. Out-of-distribution generalization under random, dense distributional shifts. arXiv preprint arXiv:2404.18370, 2024

    2024 arXiv

  21. [29]

    Y. Jin, N. Egami, and D. Rothenh \"a usler. Beyond reweighting: On the predictive role of covariate shift in effect generalization. arXiv preprint arXiv:2412.08869, 2024

    2024 arXiv

  22. [30]

    Johnson, L

    A. Johnson, L. Bulgarelli, T. Pollard, S. Horng, L. A. Celi, and R. Mark. Mimic-iv. https://physionet.org/content/mimiciv/1.0/, 2021

    2021

  23. [31]

    Kirkpatrick, R

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114 0 (13): 0 3521--3526, 2017

    2017

  24. [32]

    P. W. Koh, S. Sagawa, H. Marklund, S. M. Xie, M. Zhang, A. Balsubramani, W. Hu, M. Yasunaga, R. Phillips, I. Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning, pages 5637--5664. PMLR, 2021

    2021

  25. [33]

    Lazaridou, A

    A. Lazaridou, A. Kuncoro, E. Gribovskaya, D. Agrawal, A. Liska, T. Terzi, M. Gimenez, C. de Masson d’Autume, T. Kocisky, S. Ruder, et al. Mind the gap: Assessing temporal generalization in neural language models. In Advances in Neural Information Processing Systems, volume 34,...

    2021

  26. [34]

    J. Liu, T. Wang, P. Cui, and H. Namkoong. On the need for a language describing distribution shifts: Illustrations on tabular datasets. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36...

    2023

  27. [35]

    P. M. Long. The complexity of learning according to two models of a drifting environment. In Conference on Computational Learning Theory (COLT), 1998

    1998

  28. [36]

    Lopez-Paz and M

    D. Lopez-Paz and M. Ranzato. Gradient episodic memory for continual learning. In Advances in Neural Information Processing Systems, pages 6467--6476, 2017

    2017

  29. [37]

    Luxenberg and S

    E. Luxenberg and S. Boyd. Exponentially weighted moving models. arXiv preprint arXiv:2206.04675, 2022

    2022 arXiv

  30. [38]

    Mazzetto and E

    A. Mazzetto and E. Upfal. An adaptive algorithm for learning with unknown distribution drift. In Advances in Neural Information Processing Systems (NeurIPS), volume 36, 2023

    2023

  31. [39]

    Misra and J

    R. Misra and J. Grover. Sculpting Data for ML: The First Act of Machine Learning. Self-published, 2021. ISBN 978-0-578-83125-1

    2021

  32. [40]

    Mohri and A

    M. Mohri and A. Muñoz Medina. New analysis and algorithm for learning with drifting distributions. In International Conference on Algorithmic Learning Theory (ALT), 2012

    2012

  33. [41]

    Nasery, S

    A. Nasery, S. Thakur, and S. Sarawagi. Training for the future: A simple gradient interpolation loss to generalize along time. In Neural Information Processing Systems (NeurIPS), 2021

    2021

  34. [42]

    Nassar, J

    J. Nassar, J. Brennan, B. Evans, and K. Lowrey. BAM: Bayes with Adaptive Memory . In International Conference on Learning Representations (ICLR), 2022. URL https://openreview.net/forum?id=NdOoQnYPj_

    2022

  35. [43]

    S. J. Pan and Q. Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22 0 (10): 0 1345--1359, 2010

    2010

  36. [44]

    T.-H. Pham, X. Zhang, and P. Zhang. Non-stationary domain generalization: Theory and algorithm. In Proceedings of the Conference on Uncertainty in Artificial Intelligence (UAI), 2025

    2025

  37. [45]

    T. Qin, S. Wang, and H. Li. Generalizing to evolving domains with latent structure-aware sequential autoencoder. In International Conference on Machine Learning (ICML), 2023

    2023

  38. [46]

    Rebuffi, A

    S.-A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2001--2010, 2017

    2001

  39. [47]

    Sagawa, P

    S. Sagawa, P. W. Koh, T. B. Hashimoto, and P. Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In International Conference on Learning Representations, 2020

    2020

  40. [48]

    Sch \"o lkopf, F

    B. Sch \"o lkopf, F. Locatello, S. Bauer, N. R. Ke, N. Kalchbrenner, A. Goyal, and Y. Bengio. Toward causal representation learning. Proceedings of the IEEE, 109 0 (5): 0 612--634, 2021

    2021

  41. [49]

    Schwarz, W

    J. Schwarz, W. Czarnecki, J. Luketina, A. Grabska-Barwinska, Y. W. Teh, R. Pascanu, and R. Hadsell. Progress & compress: A scalable framework for continual learning. In International Conference on Machine Learning, pages 4528--4537, 2018

    2018

  42. [50]

    Shimodaira

    H. Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference, 90 0 (2): 0 227--244, 2000

    2000

  43. [51]

    J. Shu, Q. Xie, L. Yi, Q. Zhao, S. Zhou, Z. Xu, and D. Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In Advances in Neural Information Processing Systems, pages 1917--1928, 2019

    1917

  44. [52]

    Srebro and S

    N. Srebro and S. Roweis. Time-varying topic models using dependent dirichlet processes. Technical report, Department of Computer Science, University of Toronto, 2005

    2005

  45. [53]

    Sugiyama, S

    M. Sugiyama, S. Nakajima, H. Kashima, P. Buenau, and M. Kawanabe. Direct importance estimation with model selection and its application to covariate shift adaptation. In Advances in Neural Information Processing Systems 21 (NIPS), pages 1433--1440, 2008

    2008

  46. [54]

    Sun and K

    B. Sun and K. Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European Conference on Computer Vision, pages 443--450. Springer, 2016

    2016

  47. [55]

    N. Y. C. Taxi and L. Commission. Nyc tlc trip record data, 2025. URL https://www.nyc.gov/html/tlc/html/about/trip_record_data.shtml

    2025

  48. [56]

    Van der Vaart

    A. Van der Vaart. Asymptotic Statistics, volume 3. Cambridge university press, 2000

    2000

  49. [57]

    C. Wang, D. Blei, and D. Heckerman. Continuous time dynamic topic models. In Proceedings of the Twenty-Fourth Conference on Uncertainty in Artificial Intelligence, page 579–586. AUAI Press, 2008

    2008

  50. [58]

    Z. Wang, P. B \"u hlmann, and Z. Guo. Distributionally robust machine learning with multi-source data. arXiv preprint arXiv:2309.02211, 2023

    2023 arXiv

  51. [59]

    B. Xie, Y. Chen, J. Wang, C. Zhou, and B. Wang. Enhancing evolving domain generalization through dynamic latent representations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 15738--15746, 2024. URL https://ojs.aaai.org/index.php/AAAI/artic...

    2024

  52. [60]

    H. Yao, C. Choi, B. Cao, Y. Lee, P. W. Koh, and C. Finn. Wild-time: A benchmark of in-the-wild distribution shift over time. arXiv preprint arXiv:2211.14238, 2023

    2023 arXiv

  53. [61]

    Q. Zeng, C. Shui, B. Wang, et al. Latent trajectory learning for limited timestamps under distribution shift over time. In International Conference on Learning Representations (ICLR), 2022

    2022

  54. [62]

    Q. Zeng, L. Wang, P. Liu, C. Ling, and B. Wang. Generalizing across temporal domains with koopman operators. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17153--17161, 2024. URL https://ojs.aaai.org/index.php/AAAI/article/view/29604

    2024

  55. [63]

    Zenke, B

    F. Zenke, B. Poole, and S. Ganguli. Continual learning through synaptic intelligence. In Proceedings of the 34th International Conference on Machine Learning, pages 3987--3995. JMLR.org, 2017

    2017

  56. [64]

    Zhong and D

    X. Zhong and D. Enke. Predicting the daily return direction of the stock market using hybrid machine learning algorithms. Financial Innovation, 5 0 (1): 0 1--20, 2019

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.