Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Late Fusion Multi-task Learning for Semiparametric Inference with Nuisance Parameters

T0 review · 3 major / 5 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 Late fusion across similar tasks provably sharpens semiparametric estimates.

desk verdict The late-fusion DML package is useful and the nuisance-sharing kernel step is novel, but Theorem 3.1's faster rate does not follow from the estimator as written; the central claim needs a major fix. read the letter →

arxiv 2507.07941 v1 pith:CD76HEKI submitted 2025-07-10 stat.ME stat.ML

Sohom Bhattacharya , Yongzhuo Chen , Muxuan Liang This is my paper
classification stat.MEstat.ML MSC 62G0562G20
keywords latefusionmulti-tasklearningsemiparametricinferencenuisanceparametersdoublemachinetaskrelatednessheterogeneoustreatmenteffectsprivacy-preservingaggregationkernelregression
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 a two-step late fusion framework for multi-task learning in semiparametric models, where each task has a finite-dimensional parameter of interest $\theta_k$ and a possibly infinite-dimensional nuisance parameter $\eta_k$. The first step computes a double/debiased machine-learning estimator for each task separately; the second step aggregates these estimators through a penalized optimization that pulls similar tasks toward a common center while staying robust to outliers. The central theoretical result, Theorem 3.1, asserts that for tasks with similar parameters the aggregated estimator achieves error of order $O_P(1/\sqrt{nK})$ plus nuisance-error and bias terms, a $\sqrt{K}$-fold improvement over individual-task learning, while outlier tasks keep the individual-task rate. A second layer of late fusion for nuisance parameters, used when the nuisance functions are also similar, further improves the rates. The procedure is designed so that only gradient and Hessian summaries, not raw data, are shared across local servers, making it applicable to privacy-sensitive medical data.

What carries the argument

The central object is the aggregation objective (2.5), whose loss $\rho_k$ is a one-step Taylor expansion of each task's estimating equation around its initial estimator (2.6); this makes the objective quadratic, strictly convex, and separable across tasks, so each local server can transmit only its gradient and Hessian to the central server. The $\ell_2$ penalty $\lambda\|u_k-u_0\|_2$ toward a common center is what transfers information across tasks, with $\lambda$ controlling the bias-variance trade-off. For nuisance parameters, the machinery is a second fusion over kernel-regression losses evaluated on an $\hbar$-covering grid, which turns the infinite-dimensional estimation into many finite-dimensional fused problems. The near-orthogonality of the estimating equation keeps nuisance estimation error from dominating the parametric rate.

What would settle it

Run the proposed method on the partial-linear-model simulation of Section 4 with $K=50$ tasks all sharing the same $\theta$ and the same nuisance functions and $n=200$ per task, comparing the fused estimator's average MSE to individual-task learning over 100 repeats; the theory predicts a parametric error near $1/\sqrt{nK}$, so a failure to see a clear reduction in average MSE would contradict the central claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that the late-fusion estimator from the penalized problem (2.5)--(2.6), built from a per-task quadratic approximation to the empirical moment function plus the penalty $\lambda\|u_k-u_0\|_2$, performs no worse than the initial individual-task estimators and provably better when tasks are related. Under $(\varepsilon,\delta)$-task relatedness, Theorem 3.1 gives $\max_{k\in S}\|\widehat\theta_k-\theta_k\|_2 = O_P(1/\sqrt{nK}+\rho_n)$ plus bias terms of order $\min\{\delta,\lambda\}$ and $\varepsilon\lambda$, so the parametric component of the error shrinks by a factor $\sqrt{K}$ for tasks in the good set $S$. Theorem 3.2 applies the same fusion idea to kernel-regression estimates of infinite-dimensional nuisance functions, and Theorem 3.3 combines both layers and separates tasks into those sharing both parametric and nuisance structure, those sharing only parametric structure, and outliers, with the best rate reserved for the first group.

Load-bearing premise

The load-bearing premise is Assumption 3.1's $(\varepsilon,\delta)$-task relatedness: all but a fraction $\varepsilon$ of tasks have parameters within distance $\delta$ of one common vector $\theta_0$; if no such majority exists, the penalty cannot borrow strength without adding bias, and the claimed $\sqrt{K}$ speed-up disappears.

Editorial extensions

If this is right

  • Under the theorem's assumptions, fusing $K$ similar tasks reduces the parametric estimation error from order $n^{-1/2}$ to order $(nK)^{-1/2}$, so more tasks directly compensate for small per-task samples.
  • The method protects outliers: tasks outside the good set $S$ keep the individual-task rate $\sqrt{\log K/n}$ plus nuisance error, so pooling does not degrade a genuinely different task.
  • When nuisance functions are similar as well, the late fusion for kernel-regression nuisance estimates yields an additional improvement, giving the best overall rate for tasks in $S\cap S_\eta$.
  • The two-step pipeline shares only gradients and Hessians with a central server, so it can be deployed across hospitals or trials without sharing patient-level records.
  • In the mammography phone-consultation application, the fused estimator reports lower prediction error than individual-task learning and than the parametric transfer baseline.

Reading between the lines

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

  • The $(\varepsilon,\delta)$ relatedness assumption is not diagnosed in the paper; a reader who wanted to use the method would need a data-driven check on the spread of the initial per-task estimates to know whether the promised gain is available.
  • The privacy guarantee is operational rather than cryptographic: sharing gradients and Hessians avoids raw data transfer, but an adversarial recipient of those summaries could in principle reconstruct information about local data, a risk the paper does not analyze.
  • The nuisance-layer fusion is built specifically on kernel regression; carrying the same late-fusion idea into modern black-box nuisance learners would require new theoretical work, which the paper explicitly leaves open.
  • If tasks form several clusters rather than one central group, the same objective could be run within each cluster with a separate center $u_0$; the paper names clustering as future work rather than developing it.
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 / 5 minor

Summary. The paper proposes a late-fusion multi-task learning framework for semiparametric models with infinite-dimensional nuisance parameters. In the first step, per-task double/debiased machine-learning estimators are obtained from each data source; in the second step, these initial estimators are aggregated by solving a penalized quadratic optimization problem (2.5)/(2.8) that shrinks task-specific estimates toward a common center. A second component applies the same late-fusion idea to kernel-regression estimates of nuisance parameters. The main theoretical claims are that, for tasks whose parametric components are similar (Assumption 3.1), the fused estimator achieves the faster rate O_P(1/sqrt(nK)) up to bias terms (Theorem 3.1), and that fusion of nuisance parameters yields analogous improvements (Theorem 3.2), leading to the three-type rate classification in Theorem 3.3. The paper also presents simulations and a mammography-screening data application.

Significance. If the rate claims were correct, the paper would make a useful contribution to privacy-preserving multi-site semiparametric inference: the late-fusion architecture avoids sharing individual-level data, the explicit robustness to outlier tasks is attractive, and the extension of fusion to nuisance estimation is a genuine idea. The simulation study is reasonably extensive, covering three model classes, and the real-data application is relevant. The strengths are the clarity of the algorithmic pipeline, the explicit bias-variance trade-off in the theorems, and the effort to implement the method without raw data sharing. However, the theoretical core contains a serious internal inconsistency that affects Theorems 3.1, 3.2, and 3.3; the claimed pooled rates do not appear to be delivered by the estimator for the specified tuning parameters.

major comments (3)
  1. [Section 3.1, Theorem 3.1] The claimed 1/sqrt(nK) rate for the per-task outputs b\theta_k is not delivered by the optimization problem (2.5) when \lambda is set to C(sqrt(log K / n) + \rho_n), which is o(1). In the canonical scalar case with \rho_k(u_k) = (1/2)(u_k - e\theta_k)^2 and penalty \lambda|u_k - u_0|, the minimizer satisfies u_k = e\theta_k whenever |e\theta_k - u_0| \le \lambda, and for \lambda = o(1) the fused estimate is essentially the individual estimate e\theta_k. Even with a squared L2 penalty, u_k = (e\theta_k + 2\lambda u_0)/(1+2\lambda), so for \lambda = o(1) the variance of u_k is approximately Var(e\theta_k) = O(1/n), not O(1/(nK)). The 1/sqrt(nK) term is the rate of the common center u_0, not of the task-specific outputs b\theta_k. This is an internal inconsistency: either the theorem is about u_0, in which case the estimator output and the statement must be changed, or the tuning parameter must be O(1) so that \lambda|u_k - u_0| actually pools information, with the bias term min{\delta, \lambda} then requiring careful handling.
  2. [Section 2.2 and 2.3, equations (2.6) and (2.8)] In the running partial linear model example, the moment function is m(Z, \theta, \eta) = (Y - \mu - (T-g)\theta)(T-g), so \nabla_\theta m = -(T-g)^2 and the matrix W_k in (2.6) is negative definite. The quadratic \rho_k is then concave, and the minimization problem (2.5)/(2.8) is unbounded below for small \lambda; at best its stationary point is a maximum, not a minimum. The paper's assertion that \rho_k is 'strictly convex' assumes W_k is positive definite, which is inconsistent with the sign of the score in Example 2.1. The authors need to specify the sign convention for m so that W_k is positive definite (e.g., use the negative of the moment function) or replace the minimization with an estimating-equation/root-finding formulation.
  3. [Section 3.1, Theorem 3.2] The same tuning issue appears in the nuisance fusion theorem. The penalty is set to e\lambda = sqrt(log(K\hbar^{-2p})/(n\hbar^p)) + \hbar^2, which is of the same order as the noise of the local kernel regression estimator, whose standard deviation is about 1/sqrt(n\hbar^p). With an L1-type penalty of this size, the fused estimate b\eta_k remains close to the local kernel estimate rather than the pooled estimate, so the claimed first term sqrt(log(K\hbar^{-2p})/(nK\hbar^p)) again describes the common center rather than the per-task output. This affects the nuisance-related terms in Theorem 3.3, where ba_n^2 is used to claim faster rates from nuisance fusion.
minor comments (5)
  1. [Section 3.1, paragraph after Theorem 3.1] The paragraph says 'Before proceeding further, we discuss the implications of Theorem 3.3' but the intended reference is Theorem 3.1.
  2. [Section 2.4] The text says 'we construct a \hbar^2-covering of X' but the covering radius is not defined; please specify the relationship between the covering radius and the bandwidth \hbar, and avoid using T both for the treatment variable and for the covering set.
  3. [Section 2.4, Remark 2.3] The heading contains a typo: 'Corss-fitting' should be 'Cross-fitting'.
  4. [Section 4, first paragraph of Simulation 1] There is a typo: 'The imulation results' should be 'The simulation results'.
  5. [Section 5] When each randomized trial is split into two tasks, the two resulting datasets are not independent samples from separate populations; the paper should discuss whether this dependence affects the interpretation of the task-similarity assumption and the reported standard errors.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 3.1's rate is a conditional bound derived from stated assumptions, not a restatement of its inputs.

full rationale

The derivation is self-contained in the sense relevant to circularity. Initial estimators satisfy the stated condition (3.2), Assumptions 3.2–3.3 provide Neyman near-orthogonality and nuisance-error rates, and Assumption 3.1 defines the (epsilon, delta)-related subset S. Theorem 3.1 then states an error bound for the minimizer of (2.5) in terms of these quantities; the bound is a theorem conditional on the assumptions, not a restatement of them. The loss rho_k in (2.6) is a Taylor approximation to the estimating equation rather than a fitted surrogate, and lambda is an oracle-rate tuning parameter, not fitted to reproduce the target rate. The claimed 1/sqrt(nK) improvement may be questionable on internal-consistency grounds, since with lambda = o(1) the fused estimator does not obviously average across tasks; however, this is a correctness concern about the proof, not a circular reduction: the paper does not define theta-hat_k as the common center u0, nor does it fit lambda to the theorem's rate. The only author self-citation, Liang and Yu [31], is used as one example of an initial dimension-reduction estimator and is not load-bearing.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central rates rely on standard DML regularity (Assumptions 3.2 and 3.3), the unverifiable task-similarity Assumption 3.1, and the nuisance-similarity Assumption 3.4. The tuning parameters lambda, elambda and hbar are chosen by cross-validation; the theory only fixes their rates. No new physical or statistical entities are introduced.

free parameters (3)
  • lambda (fusion penalty for theta) = chosen by cross-validation; theory specifies lambda = C(sqrt(log K / n) + rho_n)
    Controls the strength of pulling btheta_k toward the common center u0 in (2.5). Its rate matters for the theorem, but its exact value is tuned on data.
  • elambda (fusion penalty for eta) = chosen by cross-validation; theory specifies elambda = sqrt(log(K hbar^{-2p}) / (n hbar^p)) + hbar^2
    Controls fusion of nuisance estimates in (2.9). Task-specific tuning is suggested.
  • hbar (kernel bandwidth) = chosen by task-specific cross-validation
    Bandwidth for kernel regression of nuisance functions; affects bias-variance tradeoff in Theorem 3.2.
assumptions (5)
  • domain assumption Assumption 3.1: (epsilon, delta)-task relatedness, |S^c| <= epsilon K and max_{k in S} ||theta_k - theta_0|| <= delta
    Unverifiable structural condition that a majority of tasks share a parametric component; drives the sqrt(K) rate.
  • domain assumption Assumption 3.2: Neyman near-orthogonality of m, bounded eigenvalues, Gateaux differentiability
    Restricts to moment functions where nuisance errors have second-order effect on theta; inherited from DML literature.
  • standard math Assumption 3.3: nuisance estimator rates r_n and r'_n with r'_n log^{1/2}(1/r_n) <= delta_n, and covering entropy condition
    Standard DML rate conditions.
  • domain assumption Assumption 3.4: (epsilon_eta, delta_eta)-nuisance relatedness and second-order smoothness of eta_k
    Unverifiable condition that a majority of nuisance functions are close; drives the nuisance-fusion improvement.
  • standard math Moment identification: E_k[m(Z, theta_k, eta_k)] = 0 with unique solution
    Identification of theta_k; standard in estimating equation theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Late Fusion Multi-task Learning for Semiparametric Inference with Nuisance Parameters." pith.science (2026). https://pith.science/paper/CD76HEKI

@misc{pith2026250707941,
  author       = {Pith},
  title        = {Pith review of: Late Fusion Multi-task Learning for Semiparametric Inference with Nuisance Parameters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CD76HEKI}},
  note         = {Machine review of arXiv:2507.07941}
}
read the original abstract

In the age of large and heterogeneous datasets, the integration of information from diverse sources is essential to improve parameter estimation. Multi-task learning offers a powerful approach by enabling simultaneous learning across related tasks. In this work, we introduce a late fusion framework for multi-task learning with semiparametric models that involve infinite-dimensional nuisance parameters, focusing on applications such as heterogeneous treatment effect estimation across multiple data sources, including electronic health records from different hospitals or clinical trial data. Our framework is two-step: first, initial double machine-learning estimators are obtained through individual task learning; second, these estimators are adaptively aggregated to exploit task similarities while remaining robust to task-specific differences. In particular, the framework avoids individual level data sharing, preserving privacy. Additionally, we propose a novel multi-task learning method for nuisance parameter estimation, which further enhances parameter estimation when nuisance parameters exhibit similarity across tasks. We establish theoretical guarantees for the method, demonstrating faster convergence rates compared to individual task learning when tasks share similar parametric components. Extensive simulations and real data applications complement the theoretical findings of our work while highlight the effectiveness of our framework even in moderate sample sizes.

Figures

Figures reproduced from arXiv: 2507.07941 by the authors.

Figure 1
Figure 1. Averaged and maximum MSE for Scenario 1 over 100 repeats. The x-axis represents [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Log-transformed averaged MSE for each task in Scenario 1 over 100 repeats. In each [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Averaged and Maximum MSE for the application to mammography screening study [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 41 canonical work pages

  1. [1]

    Baltruˇ saitis, T., Ahuja, C., and Morency, L.-P. (2018). Multimodal machine learning: A sur- vey and taxonomy.IEEE transactions on pattern analysis and machine intelligence, 41(2):423– 443

    work page 2018

  2. [2]

    Bastani, H. (2021). Predicting with proxies: Transfer learning in high dimension.Management Science, 67(5):2964–2984

    work page 2021

  3. [3]

    and Kohler, M

    Bauer, B. and Kohler, M. (2019). On deep learning as a remedy for the curse of dimensionality in nonparametric regression.The Annals of Statistics, 47(4):2261

    work page 2019

  4. [4]

    Bodory, H., Huber, M., and Laff´ ers, L. (2022). Evaluating (weighted) dynamic treatment effects by double machine learning.The Econometrics Journal, 25(3):628–648

    work page 2022

  5. [5]

    H., Dukes, O., and Balakrishnan, S

    Bonvini, M., Kennedy, E. H., Dukes, O., and Balakrishnan, S. (2024). Doubly-robust inference and optimality in structure-agnostic models with smoothness.arXiv preprint arXiv:2405.08525

    arXiv 2024

  6. [6]

    Cai, T. T. and Pu, H. (2024). Transfer learning for nonparametric regression: Non-asymptotic minimax analysis and adaptive procedure.arXiv preprint arXiv:2401.12272

    arXiv 2024

  7. [7]

    Chen, J., Zhou, J., and Ye, J. (2011). Integrating low-rank and group-sparse structures for robust multi-task learning. InProceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 42–50

    work page 2011

  8. [8]

    Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., and Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters: Double/debiased machine learning.The Econometrics Journal, 21(1)

    work page 2018

Show all 47 references
  1. [9]

    and Lee, Y.-Y

    Colangelo, K. and Lee, Y.-Y. (2020). Double debiased machine learning nonparametric in- ference with continuous treatments.arXiv preprint arXiv:2004.03036

    2020 arXiv

  2. [10]

    D ´ ıaz, I. (2020). Machine learning in the estimation of causal effects: targeted minimum loss-based estimation and double/debiased machine learning.Biostatistics, 21(2):353–358

    2020

  3. [11]

    Duan, R., Ning, Y., and Chen, Y. (2022). Heterogeneity-aware and communication-efficient distributed statistical inference.Biometrika, 109(1):67–83

    2022

  4. [12]

    and Wang, K

    Duan, Y. and Wang, K. (2023). Adaptive and robust multi-task learning.The Annals of Statistics, 51(5):2015–2039

    2023

  5. [13]

    Dukes, O., Avagyan, V., and Vansteelandt, S. (2020). Doubly robust tests of exposure effects under high-dimensional confounding.Biometrics, 76(4):1190–1200

    2020

  6. [14]

    and Vansteelandt, S

    Dukes, O. and Vansteelandt, S. (2021). Inference for treatment effect parameters in poten- tially misspecified high-dimensional models.Biometrika, 108(2):321–334

    2021

  7. [15]

    S., and van der Schaar, M

    Feuerriegel, S., Frauen, D., Melnychuk, V., Schweisthal, J., Hess, K., Curth, A., Bauer, S., Kilbertus, N., Kohane, I. S., and van der Schaar, M. (2024). Causal machine learning for predicting treatment outcomes.Nature Medicine, 30(4):958–968. 19

    2024

  8. [16]

    Gu, T., Han, Y., and Duan, R. (2022). Robust angle-based transfer learning in high dimen- sions.arXiv preprint arXiv:2210.12759

    2022 arXiv

  9. [17]

    and Knaus, M

    Heiler, P. and Knaus, M. C. (2021). Effect or treatment heterogeneity? policy evaluation with aggregated and disaggregated treatments.arXiv preprint arXiv:2110.01427

    2021 arXiv

  10. [18]

    J., Misra, S., and Zhang, W

    Hitsch, G. J., Misra, S., and Zhang, W. W. (2024). Heterogeneous treatment effects and optimal targeting policy evaluation.Quantitative Marketing and Economics, 22(2):115–168

    2024

  11. [19]

    Hospedales, T., Antoniou, A., Micaelli, P., and Storkey, A. (2021). Meta-learning in neu- ral networks: A survey.IEEE transactions on pattern analysis and machine intelligence, 44(9):5149–5169

    2021

  12. [20]

    Hunter, D. J. and Holmes, C. (2023). Where medical statistics meets artificial intelligence. New England Journal of Medicine, 389(13):1211–1219

    2023

  13. [21]

    Jalali, A., Ravikumar, P., and Sanghavi, S. (2013). A dirty model for multiple sparse regression.IEEE Transactions on Information Theory, 59(12):7947–7968

    2013

  14. [22]

    I., Lee, J

    Jordan, M. I., Lee, J. D., and Yang, Y. (2019). Communication-efficient distributed statis- tical inference.Journal of the American Statistical Association

    2019

  15. [23]

    Kennedy, E. H. (2023). Towards optimal doubly robust estimation of heterogeneous causal effects.Electronic Journal of Statistics, 17(2):3008–3049

    2023

  16. [24]

    and Krzy˙ zak, A

    Kohler, M. and Krzy˙ zak, A. (2005). Adaptive regression estimation with multilayer feedfor- ward neural networks.Nonparametric Statistics, 17(8):891–913

    2005

  17. [25]

    and Krzy˙ zak, A

    Kohler, M. and Krzy˙ zak, A. (2016). Nonparametric regression based on hierarchical inter- action models.IEEE Transactions on Information Theory, 63(3):1620–1630

    2016

  18. [26]

    and Langer, S

    Kohler, M. and Langer, S. (2021). On the rate of convergence of fully connected deep neural network regression estimates.The Annals of Statistics, 49(4):2231

    2021

  19. [27]

    and DiazOrdaz, K

    Kreif, N. and DiazOrdaz, K. (2019). Machine learning in policy evaluation: new tools for causal inference.arXiv preprint arXiv:1903.00402

    2019 arXiv

  20. [28]

    Lenzerini, M. (2002). Data integration: A theoretical perspective. InProceedings of the twenty-first ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, pages 233–246

    2002

  21. [29]

    T., and Li, H

    Li, S., Cai, T. T., and Li, H. (2022). 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

  22. [30]

    T., and Li, H

    Li, S., Cai, T. T., and Li, H. (2023). Transfer learning in large-scale gaussian graphical models with false discovery rate control.Journal of the American Statistical Association, 118(543):2171–2183

    2023

  23. [31]

    and Yu, M

    Liang, M. and Yu, M. (2022). A semiparametric approach to model effect modification. Journal of the American Statistical Association, 117(538):752–764

    2022

  24. [32]

    Liu, M., Zhang, Y., and Zhou, D. (2021). Double/debiased machine learning for logistic partially linear model.The Econometrics Journal, 24(3):559–588

    2021

  25. [33]

    and Zhu, L

    Ma, Y. and Zhu, L. (2012). A semiparametric approach to dimension reduction.Journal of the American Statistical Association, 107(497):168–179. 20

    2012

  26. [34]

    and Zhu, L

    Ma, Y. and Zhu, L. (2013). Efficient estimation in sufficient dimension reduction.Annals of statistics, 41(1):250

    2013

  27. [35]

    Maity, S., Sun, Y., and Banerjee, M. (2022). Minimax optimal approaches to the label shift problem in non-parametric settings.Journal of Machine Learning Research, 23(346):1–45

    2022

  28. [36]

    and Wager, S

    Nie, X. and Wager, S. (2021). Quasi-oracle estimation of heterogeneous treatment effects. Biometrika, 108(2):299–319

    2021

  29. [37]

    P., Xia, T., Watson, H

    Sanchez, P., Voisey, J. P., Xia, T., Watson, H. I., O’Neil, A. Q., and Tsaftaris, S. A. (2022). Causal machine learning for healthcare and precision medicine.Royal Society Open Science, 9(8):220638

    2022

  30. [38]

    Schmidt-Hieber, A. J. (2020). Nonparametric regression using deep neural networks with relu activation function.Annals of statistics, 48(4):1875–1897

    2020

  31. [39]

    Sidheekh, S., Tenali, P., Mathur, S., Blasch, E., Kersting, K., and Natarajan, S. (2024). Credibility-aware multi-modal fusion using probabilistic circuits.arXiv preprint arXiv:2403.03281

    2024 arXiv

  32. [40]

    and Feng, Y

    Tian, Y. and Feng, Y. (2023). Transfer learning under high-dimensional generalized linear models.Journal of the American Statistical Association, 118(544):2684–2697

    2023

  33. [41]

    Tian, Y., Gu, Y., and Feng, Y. (2023). Learning from similar linear representations: adap- tivity, minimaxity, and robustness.arXiv preprint arXiv:2303.17765

    2023 arXiv

  34. [42]

    Tripuraneni, N., Jin, C., and Jordan, M. (2021). Provable meta-learning of linear represen- tations. InInternational Conference on Machine Learning, pages 10434–10443. PMLR

    2021

  35. [43]

    and Athey, S

    Wager, S. and Athey, S. (2018). Estimation and inference of heterogeneous treatment effects using random forests.Journal of the American Statistical Association, 113(523):1228–1242

    2018

  36. [44]

    M., and Wang, D

    Weiss, K., Khoshgoftaar, T. M., and Wang, D. (2016). A survey of transfer learning.Journal of Big data, 3:1–40

    2016

  37. [45]

    Wiens, J., Guttag, J., and Horvitz, E. (2014). A study in transfer learning: leveraging data from multiple hospitals to enhance hospital-specific predictions.Journal of the American Medical Informatics Association, 21(4):699–706

    2014

  38. [46]

    and Yang, Q

    Zhang, Y. and Yang, Q. (2018). An overview of multi-task learning.National Science Review, 5(1):30–43

    2018

  39. [47]

    Zhou, D., Liu, M., Li, M., and Cai, T. (2024). Doubly robust augmented model accu- racy transfer inference with high dimensional features.Journal of the American Statistical Association, pages 1–26. 21

    2024

Pith tools

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