{"id":"7f4ebd09-891e-43ed-8a18-b7e07735c78f","arxiv_id":"2411.09056","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A thesis combining fairness-aware forecasting via non-commutative polynomial optimization with a group-blind optimal-transport bias-repair method that needs only population-level group distributions.","lead":"This machine-learning thesis proposes two fairness tools: a forecasting method with fairness constraints and a data-repair method that aligns group distributions without needing each person's race or gender. It could help organizations fix biased algorithms when collecting protected attributes is restricted, though the reported tests still lean on aggregate demographics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The without-demographics claim is untested: every experiment computes V from source data using the sensitive attribute, so no result demonstrates the framework when V is obtained externally.","rationale":"The reader's weakest_assumption is the same load-bearing concern I identified: V is computed from source data using the sensitive attribute in every experiment, so the without-demographics claim collapses if V cannot be obtained externally. My stress-test confirms that this is the most important gap. The mathematical framework itself—Theorem 5.1, Lemma 5.7, and Algorithm 1—appears internally coherent: the coupling constraint γ′V = 0 is derived correctly from the projection definition, and the Dykstra projections are standard. The issue is not a flaw in the derivation but a mismatch between the claimed operating condition (population-level V, no individual demographics) and the experimental evaluation (same-data V computed with S). The paper even acknowledges in Section 6.2 that relaxing the availability of group-wise distributions is future work, which supports the interpretation that the central claim is currently unvalidated. I therefore keep the reader's CONDITIONAL verdict: the result may be salvageable by testing with an independent V and by bounding sensitivity to noise, but as presented the headline 'without demographics' is not established.","tokens_in":55461,"tokens_out":4375,"duration_ms":49126,"concrete_test":"Re-run the Adult experiment of Section 5.4.4 with V computed entirely from an independent population-level source—e.g., a separate census table or a held-out 50% sample never used for coupling construction—and then apply the resulting coupling to the remaining data. Measure S-wise TV distance and disparate impact after repair. Additionally, run a misspecification test: perturb V by realistic noise (e.g., ±10% changes, or V from a different demographic region) and check whether parity degrades. If parity holds only when V is computed from the same source data using S, the central claim fails; if parity survives independent and noisy V, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Chapter 5 is that a coupling satisfying γ′V = 0 can be computed and applied without individual sensitive-attribute values (Theorem 5.1, Lemma 5.7, Algorithm 1). This requires V = (P^{Xs0} − P^{Xs1})/P^X to be available from population-level group distributions while the source data are unbiased samples. Section 5.4.1 states: 'since there is no such population-level information given, we directly compute V from the source data.' Computing P^{Xs0} and P^{Xs1} from source data requires the sensitive attribute of each sample, so in all reported experiments—synthetic and Adult—S is used to construct the key input; it is only omitted in the subsequent coupling/projection step. The paper's own conclusion (Section 6.2) lists as future work 'evaluating the effects when only distributions with noise are provided,' confirming that the exact external-V scenario has not been tested. Theorems 5.1 and 5.7 may be internally consistent, but without evidence that V can be obtained independently, or that parity is robust to misspecified or noisy V, the 'without demographics' claim is not validated. Moreover, in jurisdictions that bar collection of race/ethnicity even in census data—a limitation the authors themselves note—V itself is unavailable, further narrowing the practical scope. This is an experimental validation gap rather than a mathematical contradiction, but it is load-bearing because the entire contribution is defined by not requiring demographics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The thesis-style manuscript develops two optimisation frameworks for ML fairness. The first (Chapters 3–4) casts learning of linear dynamical systems as a non-commutative polynomial optimisation problem, provides convergence guarantees, and demonstrates fairness-aware forecasting on synthetic data and COMPAS. The second (Chapter 5) proposes a group-blind optimal-transport bias-repair framework. The central claim is that a coupling satisfying γ′V = 0, with V = (P^{Xs0} − P^{Xs1})/P^X, can be computed without individual sensitive-attribute values, and Algorithm 1 converges to such a coupling. The authors prove the total-repair condition in Theorem 5.1, provide KL-projection lemmas, and report experiments on synthetic data and Adult Census Income. The main weakness is that all experiments compute V directly from source data using the sensitive attribute, so the without-demographics claim is not validated in the exact scenario the framework is designed for.","tokens_in":55816,"tokens_out":3659,"duration_ms":39561,"significance":"If the Chapter 5 claim holds, it is a meaningful step: a provable demographic-parity repair method whose coupling computation and projection step do not require individual sensitive attributes, with convergence supported by Dykstra’s algorithm and Bregman projections. The manuscript provides proofs in Appendix E, reproducible code links, and clearly states limitations, including the binary-attribute restriction and the need for population-level group distributions. The total-repair guarantee is a derivation rather than a fitted prediction, and I agree with the reader’s assessment that it is not circular. The main open risk is empirical: the paper never tests the scenario where V is obtained externally, nor does it bound the effect of noisy or misspecified V. Since that scenario is the distinguishing contribution of Chapter 5, the experimental gap is load-bearing rather than cosmetic.","major_comments":[{"comment":"The without-demographics claim is not validated as stated. Section 5.4.1 explicitly says that, because no population-level information is given, V is computed directly from the source data. Computing V = (P^{Xs0} − P^{Xs1})/P^X requires the sensitive attribute S for every sample, since P^{Xs0} and P^{Xs1} are conditional distributions. Therefore all reported experiments use S to construct the key input, and S is only omitted in the subsequent coupling and projection steps. Section 6.2 confirms this gap by listing as future work 'evaluating the effects when only distributions with noise are provided.' The paper should either test the framework with V obtained from an independent population source, or clearly restrict the claim and provide an experiment with noisy or partial V. This is essential because the entire contribution is defined by not requiring demographics.","section":"Section 5.4.1, Section 6.2"},{"comment":"There is no perturbation analysis for the key quantity V. The parity guarantee γ′V = 0 is exact and relies on V coinciding with the true population-level (P^{Xs0} − P^{Xs1})/P^X. In the intended deployment scenario, V must come from a census, sandbox, or similar external source, and that source may have measurement error or may reflect a different subpopulation than the source data. The manuscript does not bound how violations of the exact equality propagate to the TV distance between projected group distributions, nor does it test V computed from an independent or noisy source. A robustness bound or a simple sensitivity experiment would materially support the central claim.","section":"Theorem 5.1, Eq. (5.16), Section 5.4.1"}],"minor_comments":[{"comment":"The notation is self-referential: 'supp(X) := {i ∈ supp(X) | V_i ≠ 0}' redefines supp(X). Use a different symbol, e.g., supp_V(X) or supp(X) ∩ {V_i ≠ 0}, to avoid confusion.","section":"Lemma 5.2"},{"comment":"Two cross-references are wrong: 'we assume X includes one neutral attribute till Section 5.3.4' should refer to Section 5.3.5 (higher dimensions), and 'The choice of target distribution will be mentioned in Section 3.6' should refer to Section 5.3.6.","section":"Section 5.3.1, Section 5.3.6"},{"comment":"The labels '1e−2-repair' and '1e−3-repair' are confusing: the text says Λ = 1e−21 and Λ = 1e−31, respectively, so the figure labels do not match the numeric values. Please use consistent notation, e.g., 10^{-21}-repair and 10^{-31}-repair.","section":"Section 5.4.4, Figure 5.6"},{"comment":"The update rule for q_{k−3} is not motivated or explained. It is presumably the standard Dykstra auxiliary-variable update, but the indexing and the distinction between k = 4,…,7 and later k are non-obvious; a short derivation or reference to the exact Dykstra variant would improve readability.","section":"Section 5.3.4, Algorithm 1"},{"comment":"The manuscript is a PhD thesis and uses thesis-style front matter (declaration, acknowledgements, publication list), while the abstract and framing say 'this paper.' If published as a journal article, the scope and framing should be adjusted accordingly.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The mathematical core of Chapter 5 appears internally consistent, and the experimental gap is fixable within the paper’s scope: add an experiment where V is obtained from an independent population source or perturbed by noise, and state the resulting scope honestly. The thesis also contains previously published material in Chapters 3–4, so the journal’s novelty assessment should concentrate on Chapter 5. I am not recommending rejection because the central mathematical claim is defensible and the missing validation is a well-defined empirical task rather than a contradiction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing worth your time here is Chapter 5. The group-blind optimal transport formulation with the constraint γ′V = 0 is genuinely new relative to Feldman, Gordaliza, and the fairness-without-demographics baselines. The derivation is clean: the parity condition is exactly the constraint, so Theorem 5.1 is a correct reduction, not a circular argument. Lemma 5.3 shows the feasible set is non-empty via the outer-product coupling, and the Dykstra-based algorithm with KL projections is supported by the cited convergence theory. This is a real contribution, and the honest statement in Section 5.5 that only one binary attribute is allowed is a fair limitation, not a hidden one.\n\nChapters 3 and 4 are re-publications of the author's own JAIR and TAC papers. They are competent consolidations, but they are not new in this thesis. The COMPAS \"state of the art\" claim in Chapter 4 is weaker than the text suggests: the error bands in Figure 4.7 overlap substantially, and the AIF360 baseline is self-selected from a toolkit rather than an independently established SOTA.\n\nThe soft spot that matters is in Chapter 5. The paper claims a bias-repair framework that does not require individual sensitive-attribute values. The algorithm itself only needs V, not S per sample. But in every experiment—synthetic and Adult—V is computed directly from the source data using the sensitive attribute, as Section 5.4.1 admits: \"since there is no such population-level information given, we directly compute V from the source data.\" That means the without-demographics claim is not actually validated as stated. The future work in Section 6.2—evaluating effects when only noisy distributions are provided—confirms that the exact external-V scenario has not been tested. This is a validation gap, not a mathematical contradiction, but it is load-bearing because the whole contribution is defined by not requiring demographics.\n\nThere is also a minor issue in Theorem 3.1: the extraction guarantee is conditioned on a flatness assumption, and the proof only cites convergence of the SDP hierarchy, so the theorem oversells what is proven. This is a minor point because the chapter is a reprint, but worth noting.\n\nWho is this for? A reader working on fairness without demographics or on OT-based repair will get value from Chapter 5's formulation and algorithm. The thesis as a whole is a solid PhD consolidation, but the new material is the OT chapter. I would give it a serious referee: the formulation deserves scrutiny and the validation gap is addressable. I would not desk-reject it.\n\nRecommendation: engage with it, but ask the authors to redo the experiments with V obtained from an independent population source, or at least with a noise-perturbed V sensitivity analysis. The central idea is worth the referee time.","headline":"The genuinely new idea is Chapter 5's group-blind OT constraint, but the central without-demographics claim is untested because every experiment computes V from source data using the sensitive attribute.","tokens_in":56306,"tokens_out":1200,"would_cite":true,"duration_ms":15478,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","90C22"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a single group-blind transport map, computed from aggregate group feature distributions alone, can repair demographic bias in a dataset, and proves convergence of an algorithm that finds such a map.","keywords":["fairness in machine learning","demographic parity","bias repair","optimal transport","group-blind projection","missing sensitive attributes","fair forecasting","non-commutative polynomial optimization"],"falsifier":"Run Algorithm 1 on a dataset where $V$ comes from independent population-level group statistics and the sensitive attribute is never used during repair; if the projected group-wise feature distributions have TV distance above the promised $\\|\\Lambda\\|_1/2$ bound, or if downstream disparate impact does not move toward 1, the central claim is refuted. A direct version on the Adult Census Income data would compare the achieved TV distance and downstream disparity between using $V$ computed from the source data and $V$ computed from a held-out population sample.","tokens_in":55243,"feed_emoji":"⚖️","tokens_out":9390,"duration_ms":87564,"temperature":0.7,"pith_summary":"The paper's central claim is that demographic bias in a dataset can be repaired without ever observing each individual's sensitive attribute, by projecting all feature vectors through one shared map. The map is derived from an optimal-transport coupling between the overall source feature distribution and a chosen target distribution, constrained so that after projection the feature distributions of the two groups coincide (total repair) or approach each other within a tunable bound (partial repair). The key identity is a vector $V$ that records how the two groups' conditional feature probabilities differ relative to the overall feature probability; the repair condition is $\\gamma' V = 0$, and a relaxed condition bounds the total-variation distance between the projected group distributions. The thesis also presents a separate framework for fairness in time-series forecasting under imbalanced data, based on non-commutative polynomial optimisation.","feed_headline":"One transport map equalizes groups without individual demographics","feed_subtitle":"Fairness repair that needs only population-level group statistics, with a provable bound on remaining disparity.","key_machinery":"The central object is the coupling $\\gamma$: an optimal-transport plan that moves probability mass from source feature values to target feature values. It is constrained by the vector $V = (P^{Xs_0} - P^{Xs_1})/P^X$, which summarises, for each feature value, how far the two groups' conditional probabilities are apart relative to the overall feature probability. Theorem 5.1 shows total repair is equivalent to $\\gamma' V = 0$; partial repair replaces this by $-\\Lambda \\le \\gamma' V \\le \\Lambda$, which bounds the total-variation distance between the projected group distributions by $\\|\\Lambda\\|_1/2$. The feasible set is the intersection of three convex sets, and Algorithm 1 uses Dykstra's algorithm with KL projections, with closed-form projections for the marginal constraints and a root-finding step for the $V$ constraint, converging to the unique coupling. Projecting source samples through the induced group-blind map then produces the repaired data.","core_discovery":"The paper introduces a bias-repair framework for transfer learning and domain adaptation in which one group-blind projection map $T$ modifies the feature values of all source samples, so that after projection the feature distributions of the two groups, defined by a binary sensitive attribute, become equal (total repair) or approximately equal with a tunable bound (partial repair). The mathematical engine is Theorem 5.1: total repair is equivalent to the coupling $\\gamma$ satisfying $\\gamma' V = 0$, where $V = (P^{Xs_0} - P^{Xs_1})/P^X$. Lemma 5.7 and Algorithm 1 give a convergent procedure, based on Dykstra's algorithm with KL projections, that solves the entropy-regularised optimal transport problem subject to this constraint, using only $V$ and the marginals $P^X$, $P^{\\tilde X}$; no individual sensitive attribute values are used to compute the coupling or to apply the projection. On the Adult Census Income dataset, the paper reports that the repaired test sets move disparate impact toward 1 and reduce the S-wise total-variation distance with little loss in f1 accuracy, whereas a barycentre baseline that does use individual demographics loses more prediction performance.","pith_inferences":["A natural next experiment the paper does not run is to supply $V$ from an independent source, such as census marginals, and measure whether demographic parity transfers to a separately collected source dataset; the current experiments compute $V$ from the source data itself.","Because the method reduces the no-demographics problem to estimating one aggregate vector $V$, its practical value hinges on how accurately $V$ can be estimated under sampling noise; perturbing $V$ and recording the resulting TV distance would quantify that sensitivity.","The framework could be extended to multiple sensitive attribute classes or to soft penalties replacing the hard constraint $\\gamma' V = 0$; the paper names the multi-class extension as future work, and a soft-penalty variant would make the fairness-distortion trade-off continuous."],"forward_implications":["If the vector $V$ is available from population-level group distributions, bias repair for demographic parity can be applied without storing or processing individual sensitive attributes, which matters where collecting such attributes is restricted or illegal.","The $\\Lambda$-relaxation gives a proven bound: after partial repair with parameter $\\Lambda$, the total-variation distance between the projected feature distributions of the two groups is at most $\\|\\Lambda\\|_1/2$, so the remaining disparity is controlled by a single tuning vector.","On the Adult Census Income dataset, the paper reports that its partial-repair schemes move disparate impact closer to 1 and reduce S-wise total-variation distance while preserving most f1 accuracy, whereas a barycentre baseline that uses individual demographics shows a larger accuracy drop.","The fair-forecasting framework in the first part of the thesis defines subgroup fairness and instantaneous fairness for time series, and the paper reports that solving the resulting min-max problems globally improves independence and separation indices relative to the original COMPAS scores."],"supporting_citations":[{"why":"Defines the total-repair and barycentre-projection framework that the paper's schemes extend to the group-blind setting, and supplies the barycentre baseline.","marker":"Gordaliza et al., 2019"},{"why":"Introduces the bias-repair idea of projecting privileged and unprivileged feature distributions to a common target; the paper's total and partial repair definitions build on it.","marker":"Feldman et al., 2015"},{"why":"Provides the convergence result for Dykstra's algorithm with Bregman projections that Lemma 5.7 uses to prove Algorithm 1 converges.","marker":"Bauschke and Lewis, 2000"},{"why":"Supplies the optimal-transport notation, the KL-projection formulation of entropic OT, and the existence of couplings used in Lemma 5.3.","marker":"Peyré et al., 2019"},{"why":"Shows how iterative Bregman projections solve entropic OT with equality constraints, the basis for the marginal projections in Algorithm 1.","marker":"Benamou et al., 2015"},{"why":"Contributes the Adult Census Income dataset on which the bias-repair and distortion experiments are run.","marker":"Becker and Kohavi, 1996"},{"why":"Provides the COMPAS benchmark and the documented recidivism-score disparity used in the fair-forecasting framework of Chapter 4.","marker":"Angwin et al., 2016"},{"why":"Defines demographic parity and fairness through awareness/unawareness, the fairness notion that total repair is designed to achieve.","marker":"Dwork et al., 2012"}],"fun_headline_variants":["Group-blind fairness: one map, no demographics needed","One transport map fixes bias without demographics","Fairness repair: one map, no demographics","Group-blind map equalizes groups, no demographics","One map repairs unfairness without individual demographics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the aggregate vector $V$, built from the two groups' feature distributions in a broader population, is available without using individual sensitive attributes and that the source data are an unbiased sample of that population; the paper's experiments compute $V$ from the source data using the sensitive attribute, so the no-demographics scenario is not tested end to end.","fun_headline_variants_meta":{"raw":{"variants":["Group-blind fairness: one map, no demographics needed","One transport map fixes bias without demographics","Fairness repair: one map, no demographics","Group-blind map equalizes groups, no demographics","One map repairs unfairness without individual demographics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000632,"raw_usage":{"total_tokens":2941,"prompt_tokens":989,"completion_tokens":1952,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":1883}},"tokens_in":605,"tokens_out":1952,"duration_ms":14592,"temperature":1.0,"reasoning_tokens":1883,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:06:46.537994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a dataset where $V$ comes from independent population-level group statistics and the sensitive attribute is never used during repair; if the projected group-wise feature distributions have TV distance above the promised $\\|\\Lambda\\|_1/2$ bound, or if downstream disparate impact does not move toward 1, the central claim is refuted. A direct version on the Adult Census Income data would compare the achieved TV distance and downstream disparity between using $V$ computed from the source data and $V$ computed from a held-out population sample.","supporting_citations":[],"review_version":1}