Pith. sign in

REVIEW 3 major objections 5 minor 33 references

FairPFN: A Tabular Foundation Model for Causal Fairness

T0 review · 3 major / 5 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 FairPFN is a tabular foundation model that, given only an observational dataset and the identity of a binary protected attribute, predicts outcomes as if that attribute's causal influence had been removed.

desk verdict Genuine new idea with extensive experiments; the biggest issue is that the implemented pretraining prior contradicts the text (binarized thresholded continuous cause, not a binary exogenous cause), which needs fixing but does not sink the method. read the letter →

arxiv 2506.07049 v1 pith:4CPKF7AG submitted 2025-06-08 cs.LG cs.CY

Jake Robertson , Noah Hollmann , Samuel Müller , Noor Awad , Frank Hutter This is my paper
classification cs.LGcs.CY
keywords causalfairnesscounterfactualtabularfoundationmodelprior-datafittednetworkprotectedattributeaveragetreatmenteffectsyntheticpriorobservationaldataonly
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

The paper presents FairPFN, a transformer pre-trained on synthetic causal-fairness data, and argues that it can identify and remove the causal effect of a binary protected attribute from tabular predictions without ever being told the causal graph. If the claim holds, a practitioner needs only the observational table and the name of the protected attribute to obtain predictions whose average treatment effect for that attribute is near zero, with predictive accuracy close to an ordinary "unfair" classifier. The authors test this on six hand-built causal scenarios and on Law School Admissions and Adult Census Income, reporting near-zero prediction ATE and counterfactual absolute error while maintaining accuracy, and they show the model correlates with fair latent noise variables similarly to a counterfactually fair baseline that is given causal information. The stated cost is a scope restriction: the protected attribute is modeled as an exogenous binary cause, and FairPFN itself reverts to a normal classifier when that assumption is violated.

What carries the argument

The load-bearing mechanism is a prior-data fitted network: a transformer pre-trained on millions of synthetic datasets sampled from a prior over sparse MLP structural causal models. For each sampled SCM, a protected attribute A is placed at an input node, draws produce a biased dataset, and a second forward pass with A's outgoing weights set to zero produces fair targets; the loss is cross-entropy against those fair targets. At inference the model integrates over this prior, approximating the posterior predictive distribution, so it never receives or outputs a graph.

What would settle it

Generate a benchmark whose protected attribute is endogenous—say group membership and outcome share a hidden confounder—with ground-truth counterfactual outcomes; if FairPFN's predictions still show near-zero average treatment effect, the exogeneity assumption is not load-bearing, and if they track the unfair classifier, the paper's own failure analysis is confirmed.

Watch

Extended reading notes

Core claim

FairPFN's central claim is that causal fairness can be achieved without any user-supplied causal model, by learning a prior over simple structural causal models and marginalizing over them at inference. During pre-training, each synthetic dataset is generated by a sparse MLP interpreted as an SCM, with a binary exogenous protected attribute A, biased observables, and a biased outcome; the same SCM is then re-run with A's outgoing edges dropped to produce a fair outcome. The transformer is trained to map the biased observational dataset to those fair outcomes, so it learns the conditional distribution $p(y_f \mid x_b, D_b)$ integrated over the prior over SCMs. On 600 synthetic datasets from six causal case studies, FairPFN is reported to sit on the fairness-accuracy Pareto front in 40% of cases and to rank 1.88 out of 4 in average-treatment-effect removal among baselines lacking causal information; on Law School and Adult it keeps prediction ATE below 0.01 and counterfactual absolute error below 0.05 and 0.2 respectively, while remaining accurate. The authors frame this as making causal fairness applicable where the causal model is unknown, at the price of assuming the protected attribute is an exogenous binary cause.

Load-bearing premise

The protected attribute must be an exogenous binary cause, and real fairness data must be drawn from causal mechanisms that lie within the family of sparse MLP structural causal models used during pre-training.

Editorial extensions

If this is right

  • A user with a tabular classification problem and a binary protected attribute can obtain causally fair predictions in a single forward pass, with no graph elicitation.
  • FairPFN's causal-effect removal improves as dataset size grows and stays stable across noise levels, so larger observational samples make the fairness guarantee more reliable.
  • Because it matches the correlation of a counterfactually fair baseline with fair noise terms, FairPFN appears to isolate fair exogenous causes rather than simply dropping the protected attribute.
  • The same network can be re-purposed as a preprocessing step that emits causally fair versions of features for downstream predictive models.
  • The approach inherits the Bayesian regularization benefits of prior-data fitted networks, which the authors cite as a reason it can beat methods optimized for group fairness on individual-level criteria.

Reading between the lines

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

  • If FairPFN's transfer holds, the practical bottleneck shifts from causal discovery to choosing which attribute is labeled 'protected': the model removes the total effect of exactly the named attribute, and the paper's own appendix shows it does not remove effects of other group attributes.
  • The proximity of FairPFN's predictions to the counterfactual-average baseline (mean difference $0.00\pm0.06$ synthetic, $0.02\pm0.04$ real) suggests Bayesian marginalization over SCMs may be closely related to averaging observational and counterfactual predictions; a formal equivalence would give a cheaper and more transparent route to causal fairness.
  • A natural stress test is endogenous or multi-category protected attributes: the paper's prior is extensible in principle, but its own failure analysis predicts a revert to a normal classifier, so the 'no causal model needed' claim is not yet a general one.
  • Because FairPFN removes the total effect, it cannot separate legally admissible spurious effects from direct and indirect discrimination; extending to path-specific removal would be needed for proportionality-style legal tests.
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. FairPFN is a tabular foundation model pre-trained on synthetic datasets generated from sparse MLP structural causal models. The pre-training objective is to predict a fair outcome obtained by removing the outgoing edges of a binary exogenous protected attribute A, given only an observational dataset and the identity of A. The paper evaluates FairPFN on six synthetic causal case studies and on the Law School Admissions and Adult Census Income datasets, using assumed causal graphs to compute ground-truth average treatment effects and counterfactual absolute errors. The reported results show that FairPFN achieves low prediction ATE and competitive accuracy relative to baselines, including baselines that use the true causal model.

Significance. If the central claim holds, FairPFN would be a meaningful contribution: it would make total-effect-elimination fairness available without a user-specified graph and would extend PFN methods to causal inference tasks. The paper provides code, a pre-trained model, and an explicit failure analysis in Appendix Figure 13, which are strengths. The significance is tempered by the fact that the prior is narrow (binary exogenous protected attribute, sparse MLP SCMs, additive Gaussian noise) and the real-world evaluation inherits the authors' assumed graphs.

major comments (3)
  1. [Section 4.1, Algorithm 2] The generated protected attribute is not the binary exogenous cause that the text claims. In Algorithm 2 the forward pass is computed from continuous exogenous values X0; only afterward is A selected from X0 and binarized over the threshold a_t, and the 'fair' forward pass zeroes row k of W0, i.e., it removes the continuous variable X0_k rather than a binary A. The pre-training prior is therefore a prior over thresholded continuous causes, not over SCMs with binary exogenous protected attributes. This breaks the identification of Y_fair as the causal effect of the binary protected attribute removed and weakens the transfer argument to real data with genuinely binary attributes. Please either generate A as a binary input to the MLP before the forward pass, or show that thresholded continuous causes are equivalent to binary causes for the pre-training objective.
  2. [Section 5.3 and Appendix A] The real-world ground-truth ATE and AE values are computed from the authors' assumed causal graphs in Figure 5 using DoWhy GCM. If these graphs are misspecified, the reported real-world fairness metrics do not measure the causal quantities the paper claims. Since the paper's headline is that no causal model is needed, the evaluation should include a sensitivity analysis over plausible graph perturbations or an additional real-world dataset with known generative mechanisms. Without this, the real-world results are conditional on the authors' modeling choices.
  3. [Section 5.2 and Appendix Figure 13] The synthetic case studies share the exogeneity assumption and additive-noise structure of the pre-training prior, and the ground-truth fairness definition is the same total-effect-elimination objective used in pre-training. The paper's own failure analysis in Figure 13 shows that FairPFN reverts to a normal classifier when the protected attribute is endogenous. The claim that FairPFN 'requires no knowledge of the causal model' should therefore be qualified: it requires that the data lie in the support of the pre-training prior over sparse MLP SCMs with exogenous protected attributes. Please add an explicit statement of this scope and a practical diagnostic for when the prior is likely to be violated.
minor comments (5)
  1. [Section 6 and Appendix C] The manuscript contains several typos, including 'perofm' (Section 6), 'optimmizing' (Section 5.3), 'post-proccessing' (Appendix C), 'mulit-category' (Appendix C), and 'Lawschool' (Figure 16).
  2. [Figure 2] The symbol ⊮ is used for the indicator function; please use \mathbb{1} for clarity.
  3. [Section 5.2 and Figure 24] The reported average ranks differ (1.88 out of 4 in the text versus 3.0 out of 7 in Figure 24); please clarify which baseline sets each rank refers to.
  4. [Appendix A] The DoWhy GCM automated search is described at a high level; please provide the exact model zoo, hyperparameters, and any random seeds used to generate the counterfactual datasets so that the real-world evaluation can be reproduced.
  5. [Section 4.1] The sentence explaining why A is sampled from {a0, a1} rather than {0,1} is unclear; the equivalence of f = 0·wx+ϵ and f = p·0x+ϵ is about the weight value, not the distribution of A.

Circularity Check

1 steps flagged · score 4.0 of 10

Algorithm 2's post-hoc binarization makes the pre-training prior a prior over thresholded proxies rather than binary exogenous causes; the synthetic prior chain is self-referential, though polynomial case studies and real-world transfer provide independent evidence.

  1. self definitional [Section 4.1, Algorithm 2 (FairPFN Synthetic Data Generation)]
    "- Sample values of exogenous causes X0 : (U × 1) ... - Select prot. attr. A, features Xbias and outcome ybias from X0, X1:H−1, and XH - Binarize A ∈ {a0, a1} over threshold at - Set input weights in row k of W0 to 0 ... - Select the fair outcome yf airfrom XH"

    The paper states that the prior models protected attributes as 'binary exogenous causes', but by construction the variable that enters the MLP and causes Xbias and Ybias is the continuous X0_k; the binary A shown to the transformer is only a thresholded observation of that continuous cause. The fair pass zeros row k of W0, which removes the effect of X0_k, not the effect of the binary A that appears in Dbias. Therefore the pre-training target Yfair is generated by removing a different variable than the protected attribute used in the dataset.

full rationale

FairPFN is not circular in the sense of fitting a parameter and then renaming that fit as a prediction: at inference it consumes only the observational dataset and the identity of the protected attribute, and the fair target is not provided. The PFN foundation is cited from prior work by overlapping authors, but that work is externally validated (TabPFN in Nature) and is used as a design basis rather than as an unverified uniqueness theorem, so self-citation is not load-bearing. The main circularity concern is in the synthetic pre-training chain: Algorithm 2 samples continuous X0, runs the MLP forward pass, then binarizes A from X0, and finally generates Yfair by zeroing row k of W0. This means the binary A is a derived thresholded proxy, while the causal edges removed are those of continuous X0_k; hence the pre-training prior is not actually a prior over SCMs with binary exogenous protected attributes as claimed. Evaluations on held-out samples from that same prior therefore measure whether the transformer learns the paper's own generation rule, which is partially self-referential. This is offset by genuine out-of-distribution evidence: the hand-crafted polynomial case studies (Figure 2) use explicit binary A structural equations, and the real-world evaluations use assumed causal graphs and external datasets; both provide independent checks that do not reduce to the pre-training objective by construction. The paper also honestly documents that FairPFN reverts to a normal classifier when the exogeneity assumption is violated (Figures 12 and 13), which limits the scope of the claim rather than making it circular. Overall, the central claim retains independent empirical content, but the Algorithm 2 construction is a definitional inconsistency in the load-bearing prior, warranting a moderate score of 4.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a strong inductive-bias assumption: the pre-training prior over sparse MLP SCMs with binary exogenous protected attributes is the right hypothesis space for real-world tabular fairness data. The paper does not report the full set of prior hyperparameters, does not ablate the prior architecture, and evaluates real-world performance against causal graphs it assumes, not ones it learns. These are pre-existing commitments the reader accepts if the result is to hold.

free parameters (1)
  • Pre-training prior hyperparameters (MLP depth, width, sparsity, noise, sample size ranges, nonlinearity set) = Not reported in paper; code referenced at github.com/jr2021/FairPFN
    These hand-chosen hyperparameters define the distribution of synthetic SCMs that the model is trained on; they shape the inductive bias and thus the claim of model-free causal fairness, but are not ablated or fully reported.
assumptions (5)
  • domain assumption Protected attributes are exogenous binary causes with no unobserved confounders.
    Assumed in pre-training (Algorithm 2), in the real-world causal graphs (Figure 5), and shown necessary by Figure 13: if A is endogenous, FairPFN reverts to a normal classifier.
  • ad hoc to paper The causal mechanisms in real-world tabular fairness data are well-approximated by sparse MLP SCMs with additive Gaussian noise sampled during pre-training.
    Defines the prior over SCMs in Section 4.1; the transfer performance hinges on this prior covering relevant real-world mechanisms, which is plausible but not demonstrated.
  • ad hoc to paper The fair outcome is defined as the outcome under intervention do(A = a) with all outgoing edges of A removed (total-effect elimination).
    This normative choice is embedded in Algorithm 1 and used as ground truth in all evaluations; it is one fairness definition among many and the paper defers direct, indirect, and spurious effect separation to future work (Section 6).
  • domain assumption The assumed causal graphs for Law School and Adult (Figure 5) are correct for evaluation purposes.
    Counterfactual datasets, ATE, and AE are computed from these graphs using DoWhy GCM (Appendix A); incorrect graphs would invalidate the fairness ground truths.
  • domain assumption In-context learning on PFNs approximates Bayesian posterior predictive inference, so integrating over the prior's SCMs is the correct inductive machinery.
    Invoked in Eq. 3 and borrowed from Hollmann et al. (2023); not proven in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FairPFN: A Tabular Foundation Model for Causal Fairness." pith.science (2026). https://pith.science/paper/4CPKF7AG

@misc{pith2026250607049,
  author       = {Pith},
  title        = {Pith review of: FairPFN: A Tabular Foundation Model for Causal Fairness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4CPKF7AG}},
  note         = {Machine review of arXiv:2506.07049}
}
read the original abstract

Machine learning (ML) systems are utilized in critical sectors, such as healthcare, law enforcement, and finance. However, these systems are often trained on historical data that contains demographic biases, leading to ML decisions that perpetuate or exacerbate existing social inequalities. Causal fairness provides a transparent, human-in-the-loop framework to mitigate algorithmic discrimination, aligning closely with legal doctrines of direct and indirect discrimination. However, current causal fairness frameworks hold a key limitation in that they assume prior knowledge of the correct causal model, restricting their applicability in complex fairness scenarios where causal models are unknown or difficult to identify. To bridge this gap, we propose FairPFN, a tabular foundation model pre-trained on synthetic causal fairness data to identify and mitigate the causal effects of protected attributes in its predictions. FairPFN's key contribution is that it requires no knowledge of the causal model and still demonstrates strong performance in identifying and removing protected causal effects across a diverse set of hand-crafted and real-world scenarios relative to robust baseline methods. FairPFN paves the way for promising future research, making causal fairness more accessible to a wider variety of complex fairness problems.

Figures

Figures reproduced from arXiv: 2506.07049 by the authors.

Figure 1
Figure 1. FairPFN Overview: FairPFN is a foundation model for causal fairness, pre-trained on synthetic datasets generated from sparse MLPs that represent SCMs with exogenous protected attributes (a). A biased dataset is created for each MLP/SCM and supplied as context to the transformer (b), with loss computed based on fair outcomes obtained by excluding the causal influence of the protected attribute (c). In practice, (d) F… view at source ↗
Figure 1
Figure 1. Biased Data Generation An MLP is randomly sampled and sparsity is induced through dropout on select edges. The protected attribute is defined as a binary exogenous variable A ∈ {a0, a1} at the input layer. We uniformly select m features X from the second hidden layer onwards to capture rich representations of exogenous causes. The target variable Y is chosen from the output layer and discretized into a binary variab… view at source ↗
Figure 2
Figure 2. Causal Case Studies: Visualization and data generating processes of synthetic causal case studies, a handcrafted set of benchmarks designed to evaluate FairPFN’s ability to remove various sources of bias in causally generated data. For each group, 100 independent datasets are sampled, varying the number of samples, the standard deviation of noise terms σ and the base causal effect wA of the protected attribute. • Av… view at source ↗
Figures from the paper (23 more)
Figure 3
Figure 3. Figure 3: Fairness Accuracy Trade-Off (Synthetic): Average Treatment Effect (ATE) of predictions, predictive error (1-AUC), and Pareto Front performance of FairPFN versus baselines in our causal case studies. Baselines which have access to causal information are indicated by a l…
Figure 4
Figure 4. Figure 4: Causal Fairness (Synthetic): Average Treatment Effect (ATE) of predictions of FairPFN compared to baselines which do not have access to causal information. FairPFN consistently removes the causal effect with a margin of error of (-0.2, 0.2) and achieves an average rank…
Figure 5
Figure 5. Figure 5: Real-World Scenarios: Assumed causal graphs of real-world datasets Law School Admissions and Adult Census Income [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Fairness-Accuracy Trade-off (Real-World): Average Treatment Effect (ATE) of predictions, predictive error (1-AUC), and Pareto Front of the performance of FairPFN compared to our baselines on each of 5 validation folds (light) and across all five folds (solid) of our re…
Figure 7
Figure 7. Figure 7: Counterfactual Fairness (Real-World): Distributions of Absolute Error (AE) between predictive distribu￾tions on observational and counterfactual datasets. Compared to baselines that do not have access to causal information, FairPFN achieves the lowest median and maximu…
Figure 8
Figure 8. Figure 8: Feature Correlation (Law School): Kendall Tau rank correlation between feature values and the predictions FairPFN compared to our baseline models. FairPFN produces predictions that correlate with fair noise terms ϵUGP A and ϵLSAT to a similar extent as the CFP baseline…
Figure 9
Figure 9. Figure 9: Effect of Base ATE (Synthetic): Distributions of prediction ATE produced by FairPFN and Unfair over quintiles (Q1-Q5) of the protected attributes’s base causal effect (base ATE). FairPFN remains consistent across quintiles, sometimes over-correcting and producing a neg…
Figure 10
Figure 10. Figure 10: Effect of Dataset Noise (Synthetic): Distributions of prediction ATE produced by FairPFN and Unfair over quintiles (Q1-Q5) of the standard deviation (std.) of exogenous noise terms in the data. FairPFN remains consistent across quintiles, while increased noise decreas…
Figure 11
Figure 11. Figure 11: Effect of Dataset Size (Synthetic): Distributions of prediction ATE produced by FairPFN over quintiles (Q1-Q5) of dataset sizes from 100-10,000 (log-scale). FairPFN becomes better at its task of removing the causal effect of protected attributes when more data is avai…
Figure 12
Figure 12. Figure 12: Multiple Protected Attributes (Synthetic): Distributions of prediction ATE and predictive accuracy produced by FairPFN vs the Unfair predictor when there are multiple protected attributes. This violates FairPFN’s prior assumptions and reverts it to a normal classifier…
Figure 13
Figure 13. Figure 13: Endogenous Protected Attributes (Synthetic): Distributions of prediction ATE and predictive accuracy produced by FairPFN vs the Unfair predictor when the protected attribute is endogenous. This violates FairPFN’s prior assumptions and reverts it to a normal classifier…
Figure 14
Figure 14. Figure 14: Graph Complexity (Prior): Distributions of Statistical Parity and predictive accuracy produced by FairPFN on prior samples with graph complexity between 10 and 200 nodes. As graph complexity increases, accuracy drops but fairness remains constant [PITH_FULL_IMAGE:fig…
Figure 15
Figure 15. Figure 15: Predictive Error (Synthetic): Predictive error (1-AUC) of FairPFN compared to our baselines. FairPFN maintains a competitive level of predictive error with traditional ML algorithms, achieving an average rank of 3.51 out of 7. 19 [PITH_FULL_IMAGE:figures/full_fig_p01…
Figure 16
Figure 16. Figure 16: Counterfactual Distributions (Law School): Predictive distributions of Unfair, Unaware, and FairPFN on observational and counterfactual versions of the Lawschool Admissions dataset. FairPFN reduces the maximum pairwise difference between these distributions to 0.05 […
Figure 17
Figure 17. Figure 17: Baseline Validation (Synthetic): Fairness-accuracy trade-off achieved by our baselines Unfair and Unaware compared to alternative choices of TabPFN (v1) and "Fairness Through Unawareness." Unfair achieves competitive performance with TabPFN (v1), while Unaware outperf…
Figure 18
Figure 18. Figure 18: Baseline Validation (Real-World): Fairness-accuracy trade-off achieved by our baselines Unfair and Unaware compared to alternative choices of TabPFN (v1) and "Fairness Through Unawareness." Our choices of baselines achieve competitive performance on the Law School Adm…
Figure 19
Figure 19. Figure 19: Aligning Counterfactual Distributions (Adult): Alignment of observational and counterfactual predictive distributions Yˆ and Yˆ a→a′ on the Adult Census Income problem. FairPFN best aligns the predictive distributions (top) and achieves the lowest mean (0.01) and maxi…
Figure 20
Figure 20. Figure 20: Statistical Parity (Synthetic): Statistical Parity (DSP) of FairPFN compared to our baselines. FairPFN achieves a similar DSP as the Random baseline and outperforms EGR which was optimized specifically for this fairness metric, achieving an average rank of 3.97 out of…
Figure 21
Figure 21. Figure 21: Group-Fairness-Accuracy Trade-off (Real-World): Statistical Parity (DSP), predictive error (1-AUC), and Pareto Front of the performance of FairPFN compared to our baselines on each of 5 validation folds (light) and across all five folds (solid) of our real-world datas…
Figure 22
Figure 22. Figure 22 [PITH_FULL_IMAGE:figures/full_fig_p023_22.png]
Figure 23
Figure 23. Figure 23: Causal Fairness Analysis (CFA) Framework: Components of the CFA framework relevant to FairPFN’s prior and evaluation. Plecko & Bareinboim (2024) Standard Fairness Model (left; SFM), which provides a meta-model for causal fairness and heavily the design of our prior, a…
Figure 24
Figure 24. Figure 24: Causal Fairness (Synthetic-All Baselines): Average Treatment Effect (ATE) of predictions of FairPFN compared to all baselines. FairPFN consistently removes the causal effect with a margin of error of (-0.2, 0.2) and achieves an average rank of 3.0 out of 7. 24 [PITH_…
Figure 25
Figure 25. Figure 25: Baseline Models: Visualization of FairPFN and our baseline models on our Fair Observable benchmark group, in terms of which variables each model is fit to and performs inference on on. Law School Admissions Adult Census Income Average Unfair 0.09±0.10 (0.00%) 0.05±0.0…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 29 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    A reductions approach to fair classification

    Agarwal, A., Beygelzimer, A., Dud \' k, M., Langford, J., and Wallach, H. A reductions approach to fair classification. In Dy, J. and Krause, A. (eds.), Proceedings of the 35th International Conference on Machine Learning (ICML'18), volume 80, pp.\ 60--69. Proceedings of Machine Learning Research, 2018

    work page 2018

  3. [3]

    Machine bias

    Angwin, J., Larson, J., Mattu, S., and Kirchner, L. Machine bias. ProPublica, May, 23 0 (2016): 0 139--159, 2016

    work page 2016

  4. [4]

    Fairness and Machine Learning: Limitations and opportunities

    Barocas, S., Hardt, M., and Narayanan, A. Fairness and Machine Learning: Limitations and opportunities. MIT Press, 2023

    work page 2023

  5. [5]

    Fair in-context learning via latent concept variables

    Bhaila, K., Van, M., Edemacu, K., Zhao, C., Chen, F., and Wu, X. Fair in-context learning via latent concept variables. 2024

    work page 2024

  6. [6]

    Causal discovery for fairness

    Binkyt \.e -Sadauskien \.e , R., Makhlouf, K., Pinz \'o n, C., Zhioua, S., and Palamidessi, C. Causal discovery for fairness. 2022

    work page 2022

  7. [7]

    G., and Cosentini, A

    Castelnovo, A., Crupi, R., Greco, G., Regoli, D., Penco, I. G., and Cosentini, A. C. A clarification of the nuances in the fairness metrics landscape. Scientific Reports, 12 0 (1), 2022

    work page 2022

  8. [8]

    and Guestrin, C

    Chen, T. and Guestrin, C. Xgboost: A scalable tree boosting system. In Krishnapuram, B., Shah, M., Smola, A., Aggarwal, C., Shen, D., and Rastogi, R. (eds.), Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD'16), pp.\ 785--794, 2016

    work page 2016

Show all 33 references
  1. [9]

    Path-specific counterfactual fairness

    Chiappa, S. Path-specific counterfactual fairness. In Hentenryck, P. V. and Zhou, Z.-H. (eds.), Proceedings of the Thirty-Third Conference on Artificial Intelligence (AAAI'19), volume 33, pp.\ 7801--7808. AAAI Press, 2019

    2019

  2. [10]

    B., Wallace, C

    Dai, H., Korb, K. B., Wallace, C. S., and Wu, X. A study of causal discovery with weak links and small samples. In Pollack, M. E. (ed.), Proceedings of the 15th International Joint Conference on Artificial Intelligence (IJCAI'95), 1997

    1997

  3. [11]

    Retiring adult: New datasets for fair machine learning

    Ding, F., Hardt, M., Miller, J., and Schmidt, L. Retiring adult: New datasets for fair machine learning. In Ranzato, M., Beygelzimer, A., Nguyen, K., Liang, P., Vaughan, J., and Dauphin, Y. (eds.), Proceedings of the 35th International Conference on Advances in Neural Informat...

    2021

  4. [12]

    and Graff, C

    Dua, D. and Graff, C. Uci machine learning repository, 2017

    2017

  5. [13]

    Equality of opportunity in supervised learning

    Hardt, M., Price, E., and Srebro, N. Equality of opportunity in supervised learning. In Lee, D., Sugiyama, M., von Luxburg, U., Guyon, I., and Garnett, R. (eds.), Proceedings of the 30th International Conference on Advances in Neural Information Processing Systems (NeurIPS'16)...

    2016

  6. [14]

    Tabpfn: A transformer that solves small tabular classification problems in a second

    Hollmann, N., M \"u ller, S., Eggensperger, K., and Hutter, F. Tabpfn: A transformer that solves small tabular classification problems in a second. In International Conference on Learning Representations (ICLR'23), 2023. Published online: iclr.cc

    2023

  7. [15]

    u ller, S., Purucker, L., Krishnakumar, A., K \

    Hollmann, N., M \"u ller, S., Purucker, L., Krishnakumar, A., K \"o rfer, M., Hoo, S. B., Schirrmeister, R. T., and Hutter, F. Accurate predictions on small data with a tabular foundation model. Nature, 637 0 (8045): 0 319--326, 2025

    2025

  8. [16]

    B., M \"u ller, S., Salinas, D., and Hutter, F

    Hoo, S. B., M \"u ller, S., Salinas, D., and Hutter, F. The tabular foundation model tabpfn outperforms specialized time series forecasting models based on simple features. 2025

    2025

  9. [17]

    O., Janzing, D., Mooij, J

    Hoyer, P. O., Janzing, D., Mooij, J. M., Peters, J., and Sch \"o lkopf, B. Nonlinear causal discovery with additive noise models. In Platt, J. and Koller, D. (eds.), Proceedings of the 22 International Conference on Advances in Neural Information Processing Systems (NeurIPS'08...

    2008

  10. [18]

    Fairness-aware classifier with prejudice remover regularizer

    Kamishima, T., Akaho, S., Asoh, H., and Sakuma, J. Fairness-aware classifier with prejudice remover regularizer. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2012, Bristol, UK, September 24-28, 2012. Proceedings, Part II 23, pp.\ 35-...

    2012

  11. [19]

    Counterfactual fairness

    Kusner, M., Loftus, J., Russell, C., and Silva, R. Counterfactual fairness. In Guyon, I., von Luxburg, U., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (eds.), Proceedings of the 31st International Conference on Advances in Neural Information Processi...

    2017

  12. [20]

    Pareto multi-task learning

    Lin, X., Zhen, H.-L., Li, Z., Zhang, Q., and Kwong, S. Pareto multi-task learning. 2019

    2019

  13. [21]

    Learning for counterfactual fairness from observational data

    Ma, J., Guo, R., Zhang, A., and Li, J. Learning for counterfactual fairness from observational data. In Singh, A. K., Sun, Y., Akoglu, L., Gunopulos, D., Yan, X., Kumar, R., Ozcan, F., and Ye, J. (eds.), Proceedings of the 29th ACM SIGKDD International Conference on Knowledge ...

    2023

  14. [22]

    Transformers can do bayesian inference

    M \"u ller, S., Hollmann, N., Arango, S., Grabocka, J., and Hutter, F. Transformers can do bayesian inference. In Proceedings of the International Conference on Learning Representations (ICLR'22), 2022. Published online: iclr.cc

    2022

  15. [23]

    Causality: Models, Reasoning and Inference

    Pearl, J. Causality: Models, Reasoning and Inference. Cambridge University Press, 2009

    2009

  16. [24]

    Causal inference on discrete data using additive noise models

    Peters, J., Janzing, D., and Sch \"o lkopf, B. Causal inference on discrete data using additive noise models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33 0 (12): 0 2436--2450, 2011

    2011

  17. [25]

    M., Janzing, D., and Sch \"o lkopf, B

    Peters, J., Mooij, J. M., Janzing, D., and Sch \"o lkopf, B. Causal discovery with continuous additive noise models. Journal of Machine Learning Research, 15: 0 2009--2053, 2014

    2009

  18. [26]

    and Bareinboim, E

    Plecko, D. and Bareinboim, E. Causal fairness analysis. Foundations and Trends in Machine Learning, 17: 0 304--589, 2024

    2024

  19. [27]

    A human-in-the-loop fairness-aware model selection framework for complex fairness objective landscapes

    Robertson, J., Schmidt, T., Hutter, F., and Awad, N. A human-in-the-loop fairness-aware model selection framework for complex fairness objective landscapes. In Das, S., Green, B. P., Varshney, K., Ganapini, M., and Renda, A. (eds.), Proceedings of the Seventh AAAI/ACM Conferen...

    2024

  20. [28]

    On causal and anticausal learning

    Sch \"o lkopf, B., Janzing, D., Peters, J., Sgouritsa, E., Zhang, K., and Mooij, J. On causal and anticausal learning. In Langford, J. and Pineau, J. (eds.), Proceedings of the 29th International Conference on Machine Learning (ICML'12). Omnipress, 2012

    2012

  21. [29]

    and Kiciman, E

    Sharma, A. and Kiciman, E. Dowhy: An end-to-end library for causal inference. arXiv:2011.04216 [stat.ME], 2020

    2011 arXiv

  22. [30]

    Counterfactual explanations without opening the black box: Automated decisions and the gdpr

    Wachter, S., Mittelstadt, B., and Russell, C. Counterfactual explanations without opening the black box: Automated decisions and the gdpr. Harvard Journal of Law and Technology, 15: 0 842--887, 2018

    2018

  23. [31]

    P., and Pechenizkiy, M

    Weerts, H., Xenidis, R., Tarissan, F., Olsen, H. P., and Pechenizkiy, M. Algorithmic unfairness through the lens of eu non-discrimination law. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pp.\ 805--816, 2023

    2023

  24. [32]

    Can fairness be automated? guidelines and opportunities for fairness-aware automl

    Weerts, H., Pfisterer, F., Feurer, M., Eggensperger, K., Bergman, E., Awad, N., Vanschoren, J., Pechenizkiy, M., Bischl, B., and Hutter, F. Can fairness be automated? guidelines and opportunities for fairness-aware automl. Journal of Artificial Intelligence Research, 79: 0 639...

    2024

  25. [33]

    Wightman, L. F. Lsac national longitudinal bar passage study. lsac research report series, 1998

    1998

Pith tools

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