{"id":"13b082c1-b182-40f4-8800-f348a3d31a72","arxiv_id":"2501.14710","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using the FiND world, a hypothetical fair world where protected attributes have no causal effect on the target, the authors show that fairness metrics become compatible and fairness aligns with accuracy, and that pre-processing methods can approximate this world in practice.","lead":"This paper argues that a hypothetical 'fair world' where protected attributes have no causal effect on outcomes dissolves both the conflict between different fairness metrics and the conflict between fairness and accuracy. It then shows that two existing data pre-processing methods can approximate this fair world, and proposes a way to check whether the approximation succeeded.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-data claim is hostage to an unvalidated causal graph: the HMDA DAG in Fig. 3 entails A⊥X_C, but the paper never tests this implication, and any violation breaks the FiND-world interpretation of the pre-processing results.","rationale":"The simulation study is internally consistent: the DAG is known by construction, so it demonstrates that if the graph is right, both pre-processing methods can approximate a FiND-like world. The HMDA experiment is the evidence for the paper's 'actionable solutions for practitioners' claim, and there the graph is assumed, not learned or validated. The DAG in Fig. 3 has race as a root independent of age/gender; this is a strong causal assumption. If it fails, the warped and adapted data need not correspond to any well-defined counterfactual world, and the observed monotonic AUC curves could arise from label equalization or distribution shift unrelated to causal fairness. This is not a criticism of the theoretical framework; the paper is explicit about the limitation in Section 6. But the conditional acceptance should hinge on validating this assumption or downgrading the real-data claim to an illustrative case study. I agree with the reader's weakest_assumption; the alternative concern about the unproven fairness-accuracy alignment in Section 2.2.4 is real, but it is partly supported by the simulation where the graph is known, whereas the DAG assumption is untested in the only real-data setting. No verdict change is needed: CONDITIONAL remains appropriate, with the condition being an explicit validation or weakening of the causal-graph assumption.","tokens_in":20517,"tokens_out":15224,"duration_ms":148399,"concrete_test":"Using the HMDA data from Section 5, test the independence A⊥X_C entailed by Fig. 3: fit a logistic regression of binary race A on the two confounders (age>62, female) and report the joint likelihood-ratio or Wald test. If the null of independence is rejected at any conventional level, the assumed DAG is falsified for this dataset, and the pre-processing results in Figs. 4 and Table 1b should be reinterpreted as conditional on a graph not supported by the data. If independence is not rejected, run the same check on the remaining implied independencies (e.g., A⊥X_P given X_C, A⊥X_D given X_C, A⊥X_A given X_C, A⊥Y given X_A, X_D, X_C) to further validate the graph.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central practical claim—that fairadapt and warping approximate the FiND world and resolve both trade-offs on HMDA—requires the assumed DAG in Fig. 3 to be correct and causally sufficient. In that DAG, race A is a root node with no edges to the confounders X_C (age, gender), so A⊥X_C is entailed. The paper does not check this or any other implication of the graph. In the actual HMDA sample, applicant age and gender distributions plausibly differ by race, which would already falsify the marginal independence and imply an unobserved common cause or a different graph. If the graph is misspecified, the pre-processing transforms (fairadapt quantile preservation, warping RPID) do not define a valid intervention to a FiND world: unremoved paths and spurious associations remain, and the increasing AUC on pre-processed data in Fig. 4 cannot be attributed to approximating the FiND world. Section 6 acknowledges the general requirement ('The assumed knowledge of the causal graph is a necessary requirement'), but an acknowledgment is not a validation; the real-data experiment is the only evidence for the 'in practice' claim, and it rests on an unverified assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that the \"FiND world\" (fictitious and normatively desired world), in which protected attributes have no causal effects on the target, resolves two classical fairness trade-offs: the incompatibility of group fairness metrics and the fairness-accuracy trade-off. The authors derive that demographic parity, equalized odds, and predictive parity all hold in the FiND world, and they propose that causal pre-processing methods (fairadapt and residual-based warping) can approximate this world. They introduce an evaluation procedure based on demographic-parity regularized training and demonstrate in simulations and on HMDA mortgage data that the pre-processed data exhibit both high fairness and high predictive performance.","tokens_in":20740,"tokens_out":14981,"duration_ms":138629,"significance":"If the claims hold, the paper offers a practically relevant recipe: pre-process data to approximate a causally defined fair world, then train unconstrained models, avoiding both the fairness-accuracy trade-off and metric incompatibility. The simulation study is a strength because the FiND world is known by construction, and the paper is transparent about the central assumption that the causal graph is known and causally sufficient. The theoretical connection between the FiND world and group fairness metrics is clearly drawn, and the use of publicly available software and data aids reproducibility. However, the real-data conclusion rests on an unvalidated causal graph, and the proposed evaluation criterion is closely tied to the very property the pre-processing methods are designed to enforce, which weakens the independence of the empirical evidence.","major_comments":[{"comment":"The assumed HMDA DAG entails A⊥X_C (race independent of the age/gender confounders), but the paper does not test this or any other empirical implication of the graph. If this independence is violated in the actual HMDA sample, the fairadapt and warping transformations do not implement a well-defined intervention to a FiND world, because unremoved paths through X_C remain, and the increasing AUC in Fig. 4 cannot be attributed to approximating the FiND world. The Section 6 acknowledgment that a correct DAG is a necessary requirement is not a substitute for empirical validation; please test the DAG's implied marginal independences on the HMDA data, or provide a sensitivity analysis under alternative plausible DAGs.","section":"Section 5, Fig. 3"},{"comment":"The proposed method for evaluating whether pre-processing approximates the FiND world is based on the same property (demographic parity holds in the FiND world) that the pre-processing methods are specifically designed to produce. Observing that DP-regularized models achieve higher AUC on pre-processed test data is therefore close to a consistency check rather than an independent validation of the FiND approximation. To make the criterion informative, the authors should demonstrate its discriminative power, for example by including a negative control (such as a pre-processing method that removes only some causal paths) and showing that the proposed criterion fails to indicate FiND approximation in that case, or by stating precisely what outcome would count as a failure.","section":"Section 4.1, Algorithm 1"},{"comment":"The paper claims to \"show theoretically\" that fairness aligns with high predictive performance in the FiND world, but Section 2.2.4 provides only a heuristic argument rather than a formal theorem or derivation. In particular, no proof is given that a model trained on biased real-world data under a DP constraint will have higher AUC on FiND-world test data; this is demonstrated only in a specific simulation. Since this alignment is a central component of the paper's claim to resolve the fairness-accuracy trade-off, please either supply a precise theorem with explicit conditions, or revise the wording so that the theoretical contribution is limited to the derivation that the FiND world satisfies the group fairness metrics, with the alignment claim presented as an empirical finding.","section":"Section 2.2.4"}],"minor_comments":[{"comment":"In the real-world row, the 95% confidence interval for AUC is reported as [0.895,0.899], which excludes the point estimate 0.887; this is internally inconsistent and should be corrected.","section":"Table 3"},{"comment":"The FiND row contains the typo \"unkown\"; it should read \"unknown\".","section":"Table 1b"},{"comment":"The graphoid argument for equalized odds and predictive parity is correct only because the FiND DAG makes A d-separated from all other variables, giving joint independence A⊥(Y*,Ŷ*). The text should state this d-separation explicitly, since readers may otherwise reasonably object that marginal independence of A from Y and from Ŷ does not generally imply conditional independence given Y or Ŷ.","section":"Section 2.2.1"},{"comment":"The sentence \"the performance of the real world model is not directly interpretable as it still inherits bias\" is vague; please specify which models in Table 1a are directly comparable and in what sense the real-world AUC is not interpretable.","section":"Section 4.3"},{"comment":"The x-axis is described in the text as fairness 1−C(π̂_test), but the figure labels are not described in the caption; please clarify the axis definitions in the caption for all panels.","section":"Figure 2"},{"comment":"The phrase \"unbiased data\" is ambiguous because, in a statistical sense, unbiasedness refers to estimators; here it means data generated from the FiND world. Please rephrase to avoid confusion.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript builds substantially on the authors' own prior work: the FiND world [9] and the warping method [8] are introduced by the same group. While they are cited appropriately, the paper should more explicitly delineate which of the theoretical results are new relative to [9] and which elements of the practical pipeline are inherited. The sole real-data validation (HMDA) rests on a causal graph whose testable implications are not checked; this is the main risk to the 'in practice' claim. I would encourage the editor to require the proposed DAG validation or a robustness analysis before acceptance. The paper is otherwise clearly written and the simulation study is a useful proof of concept."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the genuinely useful part is the in-processing diagnostic and the empirical demonstration that fairadapt and warping can approximate the FiND world in simulation and on HMDA. The theoretical component is a repackaging of known independence facts, and the paper stumbles on its own derivation in Section 2.2.1.\n\nWhat it does well: it takes the FiND world idea from the authors' earlier work and gives it a concrete evaluation scheme. The simulation study is the right design: you know the true counterfactual world, so you can actually check whether the pre-processed data match it. The HMDA experiment is a reasonable stress test, and the paper is transparent about the causal graph requirement in Section 6.\n\nWhere it's soft. First, the claim that A⊥Y and A⊥Ŷ imply Ŷ⊥A|Y using 'compositional and weak union axioms' is wrong as stated; marginal independence does not give conditional independence. The conclusion is salvageable because in the FiND world A is independent of the whole vector (X*,Y*), so joint independence holds, but the text needs to say that. Second, the success criterion for pre-processing is theory-laden. You train with a DP penalty and check whether AUC on pre-processed data rises; this is exactly the behavior the paper predicts for the FiND world. In simulation that is backed by ground truth; on HMDA it is the only evidence, so the inference is weaker than it looks. Third, the HMDA DAG is assumed, not validated. Figure 3 entails A⊥X_C (race independent of age/gender), which is implausible for this data, and the paper never checks it. If that edge exists, the pre-processing does not define a valid intervention to a FiND world, and the AUC increase in Figure 4 cannot be attributed to FiND-world approximation. Section 6 acknowledges the need for a correct graph, but that is not the same as testing it.\n\nOverall, this is a serious, well-written paper that deserves referee time. I would want the derivation fixed, a robustness check on the HMDA graph (at least a test of A⊥X_C), and a sensitivity analysis to graph misspecification before accepting. But it is a contribution worth engaging with.","headline":"The simulation diagnostic is the real contribution; the theory is familiar and the HMDA causal graph is unvalidated.","tokens_in":21293,"tokens_out":5345,"would_cite":true,"duration_ms":42587,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Two causal pre-processing methods resolve both fairness trade-offs by approximating the FiND world.","keywords":["causal fairness","FiND world","fairness-accuracy trade-off","impossibility theorem","causal pre-processing","fairadapt","residual-based warping","demographic parity"],"falsifier":"On a simulated FiND world dataset of the paper's type, train the same gradient-boosted model at the 11 penalty strengths of Algorithm 1 and compute the Spearman correlation between the demographic-parity violation $C$ and AUC on the FiND world test data. The paper's claim predicts a negative correlation (fairer models are more accurate); a positive or zero correlation would refute the alignment claim.","tokens_in":20273,"feed_emoji":"⚖️","tokens_out":6160,"duration_ms":52488,"temperature":0.7,"pith_summary":"This paper argues that the two central deadlocks of fair machine learning—the fairness-accuracy trade-off and the impossibility of satisfying several group fairness metrics at once—are artifacts of evaluating models on biased, real-world data. In the 'fictitious and normatively desired' (FiND) world, where the protected attribute has no causal effect on the target, either directly or indirectly, demographic parity, equalized odds, and predictive parity all hold by construction, and better fairness corresponds to better predictive performance. The authors show that two causal pre-processing methods, fairadapt and residual-based warping, approximate the FiND world well enough on simulated and real mortgage data that both trade-offs disappear in practice. This matters because it offers practitioners a concrete route to fair models without sacrificing accuracy and without choosing among competing fairness metrics.","feed_headline":"Causal 'fair world' dissolves both AI fairness trade-offs","feed_subtitle":"Disconnect protected attributes from outcomes, and accuracy and the main group fairness metrics all align at once.","key_machinery":"The central object is the FiND (fictitious and normatively desired) world, a counterfactual causal model in which the protected attribute is a root node with no children, so no causal path from it reaches any feature or the target. Its work in the argument is to instantiate the equal-base-rate exception to the impossibility theorem through the independence $Y \\perp\\!\\!\\perp A$ and, at the same time, to turn the fairness-accuracy trade-off into an alignment. Two pre-processing mechanisms carry the practical half: fairadapt, which builds 'fair twins' by quantile-preserving projection onto a single baseline protected group, and residual-based warping, which computes rank-preserving interventional distributions so protected-group individuals keep their ranks while their values are moved to the unprotected group's distribution. The evaluation mechanism is a regularized empirical risk that adds a demographic-parity penalty term to gradient-boosted-tree training and traces AUC against penalty strength, providing the empirical signature of FiND-world approximation.","core_discovery":"On the paper's own terms, the central claim is that the FiND world—a counterfactual world in which the protected attribute has no causal effect on the target, neither directly nor indirectly—provides the common cause behind earlier empirical findings that fairness and accuracy can align on unbiased data. Because the protected attribute is independent of the target in this world, group base rates are equal, and the FiND world becomes the equal-prevalence special case under which the impossibility theorem no longer bites: demographic parity, equalized odds, and predictive parity are satisfied simultaneously. The same independence makes fairness and accuracy positively related: models trained in the real world that enforce a fairness constraint perform better when evaluated in the FiND world. The paper further claims that pre-processing methods that approximate the FiND world transfer this theoretical resolution to practice, and the authors provide an in-processing criterion—whether increasing the strength of a demographic-parity constraint improves AUC on pre-processed test data—for judging when the approximation succeeds.","pith_inferences":["If the causal graph is learned or misspecified, the methods' success should degrade in a predictable way; a sensitivity analysis over plausible DAG perturbations would turn this assumption into a measured quantity.","The FiND-world argument suggests that any pre-processing method that achieves the same interventional independence—not only causal ones—should also dissolve the trade-offs, so comparing non-causal fair-representation learners on the same evaluation scheme is a direct testable extension.","Extending to multiple protected attributes and intersectional groups should inherit the same base-rate-equality mechanism only if the graph blocks all causal paths from every protected attribute, making intersectionality a graph-sufficiency question rather than a metric-choice question.","Since the FiND world equalizes base rates by normative fiat, the practical guarantee matches the normative intent only when historical discrimination is in fact the sole source of base-rate differences; where group differences have legitimate causes, the approximation silently removes them too."],"forward_implications":["In the FiND world, demographic parity, equalized odds, and predictive parity all hold simultaneously without imposing constraints, so the impossibility theorem loses force whenever the data represent that world.","On data approximating the FiND world, increasing the strength of a fairness constraint during training no longer costs AUC; the fairness-performance curve slopes upward instead of downward.","Practitioners can stop explicitly enforcing a chosen fairness metric and instead train unconstrained models on pre-processed data, because the approximation already guarantees group-level fairness.","Both fairadapt and warping achieve near-complete group fairness (differences under about 3.6% in simulation) while keeping AUC essentially unchanged, and the real HMDA mortgage data shows the same pattern.","The paper's evaluation method gives a general test: if enforcing a demographic-parity constraint raises AUC on pre-processed test data, the pre-processing has successfully approximated the FiND world."],"supporting_citations":[{"why":"Defines the FiND world (fictitious and normatively desired world) and its philosophical grounding, which the paper's theoretical resolution of both trade-offs is built on.","marker":"[9]"},{"why":"Introduces the fairadapt pre-processing method with quantile preservation; one of the two causal methods used to approximate the FiND world.","marker":"[50]"},{"why":"Introduces the residual-based warping method based on rank-preserving interventional distributions, the second causal pre-processing method evaluated.","marker":"[8]"},{"why":"Shows that fairness constraints can improve accuracy when data are biased, the empirical phenomenon the paper re-derives causally in the FiND world.","marker":"[7]"},{"why":"Shows that the fairness-accuracy trade-off is an artifact when accuracy is measured on fair (unbiased) labels, the result the paper explains and operationalizes through pre-processing.","marker":"[59]"},{"why":"States the impossibility theorem that equalized odds and predictive parity cannot hold simultaneously except in special cases; the FiND world realizes the equal-base-rates exception.","marker":"[15]"},{"why":"Gives the complementary impossibility result for risk scores, which the paper's equal-base-rate argument also bypasses in the FiND world.","marker":"[35]"},{"why":"Characterizes the exceptions to the impossibility theorem and 'fairness regions,' providing the theoretical context for why equal base rates in the FiND world resolve metric incompatibility.","marker":"[4]"}],"fun_headline_variants":["Fair world removes fairness-accuracy and metric conflicts","Causal preprocessing aligns fairness metrics and boosts accuracy","Cut protected attribute links to make AI both fair and accurate","Unbiased world makes fairness metrics and accuracy agree","Remove causal effect of protected attribute for fair and accurate AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The causal DAG supplied by the practitioner must match the true data-generating process, including no hidden confounders; if the graph is misspecified, unremoved causal paths survive and the pre-processing no longer approximates the FiND world, as the paper concedes in its discussion.","fun_headline_variants_meta":{"raw":{"variants":["Fair world removes fairness-accuracy and metric conflicts","Causal preprocessing aligns fairness metrics and boosts accuracy","Cut protected attribute links to make AI both fair and accurate","Unbiased world makes fairness metrics and accuracy agree","Remove causal effect of protected attribute for fair and accurate AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001241,"raw_usage":{"total_tokens":5118,"prompt_tokens":997,"completion_tokens":4121,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":4046}},"tokens_in":613,"tokens_out":4121,"duration_ms":30731,"temperature":1.0,"reasoning_tokens":4046,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:52:33.800089+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a simulated FiND world dataset of the paper's type, train the same gradient-boosted model at the 11 penalty strengths of Algorithm 1 and compute the Spearman correlation between the demographic-parity violation $C$ and AUC on the FiND world test data. The paper's claim predicts a negative correlation (fairer models are more accurate); a positive or zero correlation would refute the alignment claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the fairadapt pre-processing method with quantile preservation; one of the two causal methods used to approximate the FiND world."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the residual-based warping method based on rank-preserving interventional distributions, the second causal pre-processing method evaluated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that fairness constraints can improve accuracy when data are biased, the empirical phenomenon the paper re-derives causally in the FiND world."},{"cited_title":"Wick, Swetasudha Panda, and Jean-Baptiste Tristan","cited_arxiv_id":null,"evidence_quote":"Shows that the fairness-accuracy trade-off is an artifact when accuracy is measured on fair (unbiased) labels, the result the paper explains and operationalizes through pre-processing."}],"review_version":1}