{"id":"1b3968eb-5e27-4e3f-9ad2-eca6bf3d2796","arxiv_id":"2412.03009","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DataSift orders data-pool points by predicted effect on demographic parity and uses a bandit over partitions to acquire them, improving fairness with less data than baselines.","lead":"This paper proposes DataSift, a system that uses multi-armed bandits and influence functions to choose which new data points to acquire so a model becomes fairer faster. The authors report that buying a small fraction of a data pool can reduce demographic parity gaps more effectively than random or accuracy-focused acquisition.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (1) and Algorithm 1 define rewards with signed ΔF while accepting batches by absolute disparity, so improvements from a positive initial F or overshoots are scored as penalties; the UCB score is also undefined for arms with no positive reward.","rationale":"The reader's conditional verdict already identifies a sign inconsistency in the reward formula and an undefined UCB term in the rationale, so my read does not move the verdict. However, the reader's formally stated weakest_assumption concerned the accuracy of first-order influence functions and the transfer of the ridge regressor in DataSift-Inf; my primary concern is more basic and affects DataSift itself: the reward/UCB update as written would penalize the very acquisitions the algorithm claims to reward in any setting where F is positive or crosses zero. This is an internal inconsistency, not a disagreement with prior consensus, and it can be settled by a small synthetic experiment or by inspecting the released code. I do not treat the availability of a source-code link as sufficient verification, because the manuscript's formulas and the code may diverge, and the paper does not state which one produced Figures 3-8. If the code uses a corrected reward, the text must be updated; if it uses the printed formula, the empirical results need re-examination. Either way, the conditional verdict stands until the inconsistency is resolved and the experiments are re-run or justified.","tokens_in":19725,"tokens_out":7308,"duration_ms":79346,"concrete_test":"Construct a synthetic two-partition data pool with an initial logistic-regression model that has F = +0.2 on the test set, and engineer one partition so that its points move F toward +0.05 while the other partition's points leave F unchanged. Run Algorithm 1 exactly as printed with Eq. (1) and the stated definition of n_i^k. If the improving partition receives a negative reward and the algorithm fails to reduce disparity, the published reward rule is incompatible with the claimed behavior; repeat with the corrected reward −(|F_new| − |F_old|) to confirm the intended mechanism works.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The written MAB update cannot be the mechanism that produced the reported fairness gains. In Section 3.2, fairness is defined as F_D = P(Ŷ=1|S=0) − P(Ŷ=1|S=1), and lower |F_D| means higher fairness. Algorithm 1 line 12 computes ΔF = F(D∪b) − F(D), and line 13 retains a batch only when |F(D∪b)| ≤ |F_best|. But Eq. (1) then rewards partition C_j with r_j = ΔF / ((1+|ΔBR_Cj|)(1+dist(C_i,C_j))), explicitly stating that ΔF > 0 is a reward and ΔF < 0 is a penalty. Whenever the current model has F > 0, an improvement toward zero gives ΔF < 0 and is penalized; the same happens if a good acquisition overshoots across zero. Thus the partition responsible for improving fairness is pushed down by UCB instead of up. The problem is compounded by the definition of n_i^k as the number of times C_i is 'selected and rewarded a positive score': any arm receiving a penalty or never yet rewarded has n_i^k = 0, making R_i^k and U_i^k in Eq. (2) undefined. Since Section 4.4 shows that the reward-score choice is what separates DataSift from AutoData, this sign error directly threatens the central claim that DataSift identifies the right partitions. The paper must either correct Eq. (1) to use −(|F_new| − |F_old|) (or an equivalent signed improvement), define n_i^k as a selection count, and re-run the experiments, or explicitly state and justify that all reported runs have F < 0 throughout so that the printed formula happens to coincide with improvement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the problem of acquiring additional labeled data to improve the demographic parity fairness of a downstream classifier. It proposes DATASIFT, an MAB-based approach that partitions the data pool, treats partitions as arms, and uses a UCB policy with a fairness-aware reward that also incorporates base-rate difference and inter-partition distance. A second method, DATASIFT-Inf, augments this with influence functions to rank data points within a partition and build batches without retraining. The authors empirically compare both methods against Random, Entropy, Inf, and AutoData on six real-world datasets and several classifiers, reporting that DATASIFT improves fairness faster and with less budget than baselines, and that DATASIFT-Inf further accelerates convergence.","tokens_in":20141,"tokens_out":6310,"duration_ms":55252,"significance":"If the results hold, the paper addresses a timely and underexplored problem: how to prioritize additional data acquisition to mitigate group fairness violations, rather than applying post-hoc bias mitigation. The MAB formulation and the integration of influence functions for fairness-oriented data valuation are reasonable and could be of practical value. The authors provide a link to source code, which is a positive step for reproducibility. However, the paper's central mechanism contains a sign inconsistency between the reward and the acceptance criterion, the UCB update is undefined for arms without positive rewards, and the empirical evaluation lacks variance information. These issues currently undermine the validity of the reported gains and need to be resolved before the contribution can be assessed reliably.","major_comments":[{"comment":"The reward definition in Eq. (1) is inconsistent with the acceptance criterion in Algorithm 1. The algorithm retains a batch when |F(D^train ∪ b)| ≤ |F_best| (line 13), i.e., when the absolute demographic parity gap decreases. However, Eq. (1) rewards partitions with r_j = ΔF / ((1+|ΔBR_Cj|)(1+dist(C_i,C_j))) where ΔF = F_{new} − F_{old}, and the text states that ΔF > 0 is a reward and ΔF < 0 is a penalty. When the current model has F > 0 (bias against the privileged group), moving toward zero gives ΔF < 0, so the very partitions that improve fairness are penalized; the same occurs when an acquisition overshoots across zero. Since Figure 5 plots signed F values (negative and positive), the precondition F < 0 throughout is not met. This sign error means the described UCB mechanism would push the search away from useful partitions, and the reported fairness gains cannot be produced by the algorithm as written. The paper must correct Eq. (1) to reward signed absolute improvement, e.g., r_j = −(|F_new| − |F_old|)/... or an equivalent, and re-run all experiments.","section":"3.2, Eq. (1) and Algorithm 1 lines 12–13"},{"comment":"The definition of n_i^k as the number of times C_i is 'selected and rewarded a positive score' breaks the UCB update. R_i^k = (1/n_i^k) Σ r_j^i is undefined for any arm that has never received a positive reward (n_i^k = 0). Moreover, if no arm has a positive reward after some iterations, n^k = Σ_i n_i^k = 0, making ln(n^k) in Eq. (2) undefined. This is not a corner case: the sign error in Eq. (1) makes negative rewards common, so the algorithm cannot compute U_i^k for the arms it most needs to explore. The paper should redefine n_i^k as the number of times the arm was selected (regardless of reward), and define R_i^k as the average over all observed rewards (or use a standard UCB with a prior).","section":"3.2, definition of n_i^k and Eq. (2)"},{"comment":"The central empirical claim that DATASIFT 'consistently outperforms' baselines is supported only by single-run curves. The methods involve random batch sampling (DataSift), stochastic partition selection (GMM), and MAB exploration, so the reported trajectories are subject to high variance. Without error bars, standard deviations over multiple seeds, or significance tests, the reader cannot determine whether the observed advantages are real or artifacts of one run. Please report means and standard deviations over at least 5–10 random seeds (or equivalent) for all main figures, and perform a paired significance test where appropriate.","section":"4, Figures 3–6 and Table 2"}],"minor_comments":[{"comment":"There is a typo: 'Alorithm 2' should be 'Algorithm 2'.","section":"4.3"},{"comment":"'know to be NP-hard' should be 'known to be NP-hard'.","section":"3.3.2"},{"comment":"'to acquire mode data points' should be 'to acquire more data points'.","section":"3.2"},{"comment":"The y-axis label is 'Fairness Metric' in both, but Figure 3 uses absolute disparity (non-negative) whereas Figure 5 shows signed F values (including negative numbers). Please use a consistent definition, e.g., plot |F| or label the signed quantity clearly.","section":"Figures 3 and 5"},{"comment":"Algorithm 2 line 3 says 'Train Regressor R' but does not specify the model class; later text refers to a ridge regressor. Please state the regressor in the algorithm.","section":"Algorithm 2 and Section 3.3.2"},{"comment":"The source code link appears as the word 'DATASIFT' without a URL; include a working hyperlink or footnote.","section":"4.1.3"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eq. (1) is a load-bearing flaw: as written, the reward does not match the fairness-improvement criterion, so the reported experimental outcomes cannot be attributed to the described algorithm. The n_i^k issue compounds this by making the UCB update undefined for arms that only receive penalties. Both are fixable, but the authors must correct the definitions, re-run the experiments, and provide variance information before the claim of consistent outperformance can be trusted. I would not reject on the basis of the influence-function heuristic alone, because the authors explicitly acknowledge that the selected batch is not guaranteed to be optimal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the combination of MAB and influence functions for fairness-driven data acquisition is a real extension, and the empirical study is fairly broad. But as written, the reward formula in Eq. (1) has the sign backwards relative to the algorithm's own acceptance rule, and the UCB exploration term can go imaginary for arms that are always rewarded. That means the printed algorithm is not the one that could have produced the reported gains. Fixable, but needs to be fixed and the experiments re-run.\n\nWhat's new: AutoData and Slice Tuner already do acquisition for fairness/accuracy, but the base-rate-adjusted reward that spreads credit across partitions with similar protected-group composition is a sensible twist, and the influence-function regressor to rank unlabeled points within a partition is a useful speedup. The paper does a service by framing data acquisition as a fairness lever and testing on six real datasets plus multiple classifiers.\n\nSoft spots: The sign issue is load-bearing. With F defined as P(ŷ=1|S=0) − P(ŷ=1|S=1), an improvement from positive F moves ΔF negative, so Eq. (1) would penalize the very partition that helped. Also, n_i^k is defined as selections with positive reward, so an arm with all-negative rewards has n_i^k=0, and Eq. (2) becomes undefined; if an arm is always rewarded, the sqrt goes imaginary. These aren't cosmetic. The experiments also lack error bars or multiple seeds, and there's no comparison to Slice Tuner, the most directly relevant prior work. The influence-function transfer is heuristic, but they acknowledge that. The accuracy table shows some big drops for the Inf baseline, which is fine, but highlights the need for seed variance.\n\nWho it's for: researchers working on data-centric fairness and data acquisition. It deserves a serious referee, but the authors need to correct the reward sign and the UCB count, re-run the experiments, and add the missing baseline before it's trustworthy.\n\nRecommendation: Send to peer review, but with a clear request for major revision. The idea is worth engaging with.","headline":"Solid idea, but the written algorithm has a reward-sign bug that undermines the central claim; fixable and worth a look.","tokens_in":20628,"tokens_out":3929,"would_cite":false,"duration_ms":36105,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a multi-armed bandit that ranks candidate data points by their predicted effect on demographic parity can shrink a model's fairness gap much faster than random, entropy, or accuracy-focused acquisition.","keywords":["data acquisition","algorithmic fairness","demographic parity","multi-armed bandits","upper confidence bound","influence functions","data valuation","reinforcement learning"],"falsifier":"Take any of the six datasets and a random sample of unseen pool points; for each, retrain once with that single point added, record the actual change in demographic parity, and correlate those changes with the influences predicted by the ridge regressor. A near-zero or negative correlation would falsify the valuation layer, and then DataSift-Inf would reduce to random batch selection.","tokens_in":19518,"feed_emoji":"⚖️","tokens_out":5952,"duration_ms":55913,"temperature":0.7,"pith_summary":"This paper aims to show that model fairness can be improved at the data-procurement stage: instead of buying arbitrary or balanced extra labels, an algorithm can decide which specific points from a pool are worth acquiring. It proposes DataSift, which splits the pool into partitions, treats each partition as an arm in a multi-armed bandit, and uses a reward score that rewards partitions whose data improves demographic parity while preserving accuracy. A faster variant, DataSift-Inf, ranks points inside the chosen partition by the influence functions predict each would have on fairness, so each batch is deliberately constructed rather than random. Across six datasets, the authors report that even a small acquisition budget reduces the demographic-parity gap substantially, often matching or beating acquisition of the entire pool.","feed_headline":"A bandit picks which labels to buy to shrink AI bias","feed_subtitle":"Ranking extra training points by predicted fairness impact closes most of the demographic-parity gap with a fraction of the budget.","key_machinery":"The load-bearing pieces are the UCB bandit over partitions and the influence-function valuation layer. For DataSift, Equation (1) is the reward-spreading rule: when partition $C_i$ is evaluated, partition $C_j$ receives score $r_j = \\Delta F / ((1 + |\\Delta BR_{C_j}|)(1 + \\mathrm{dist}(C_i,C_j)))$, where $\\Delta F$ is the observed demographic-parity change, $\\Delta BR$ is the intra-partition base-rate difference, and $\\mathrm{dist}$ is normalized centroid distance. For DataSift-Inf, the valuation machinery is the first-order influence function $I_\\theta(d) = -H_\\theta^{-1} \\nabla_\\theta L(d,\\theta^*)$, applied to fairness $F$ through the chain rule, then transferred to unlabeled pool points by a ridge regressor $R$ trained on (training point, predicted influence) pairs. The regressor's output orders each partition, and the top-$K$ points form each batch.","core_discovery":"On the authors' own terms, the central discovery is that fairness is a property that can be bought selectively: the improvement in demographic parity depends less on how many points are added than on which points are added, and the right ordering can be learned. DataSift formalizes acquisition as a UCB bandit problem where the arms are data-pool partitions and the reward for a partition is the measured fairness improvement of a batch, shared with other partitions according to their base-rate differences and centroid distances. DataSift-Inf adds a valuation layer: first-order influence functions computed on the current training model estimate each training point's effect on the fairness metric, a ridge regressor learns to predict those influences for points in the pool, and the top-ranked points in the selected partition form the next batch. The reported result is that DataSift-Inf reaches the fairness threshold with less than half the budget in most datasets and stops early, while the plain bandit already outperforms random, entropy, and accuracy-oriented baselines.","pith_inferences":["The valuation-plus-bandit design is not tied to demographic parity: any fairness signal that can be computed after each batch, or differentiated where needed, could be plugged into Equation (1), so equalized odds or calibration gaps are natural next targets.","Because DataSift-Inf's regressor is trained on the current model's influence values, its reliability should degrade as the model moves far from the point where those influences were computed; a direct test would measure correlation between predicted and actual per-point fairness effects on held-out pool points.","A fixed early-stopping threshold caused oscillation in at least one dataset; a dataset-adaptive threshold derived from the initial disparity would make the budget-to-fairness trade-off more predictable."],"forward_implications":["Acquiring a small, well-chosen set of additional labels can close most of the demographic-parity gap, in some cases matching or beating a model trained on the entire data pool.","Balanced acquisition across demographic groups is not a reliable fairness strategy; selective ordering is what carries the benefit.","DataSift-Inf can stop early once fairness crosses a threshold, so the labeling budget and retraining cost both shrink.","The approach preserves or improves accuracy on most tested datasets even though it optimizes for fairness, with slight drops on two datasets that a larger budget can offset."],"supporting_citations":[{"why":"Defines MAB-based selective data acquisition for model accuracy; DataSift adapts this arm-selection idea to fairness and uses it as the AutoData baseline.","marker":"[15]"},{"why":"Supplies the first-order influence functions that DataSift-Inf uses to estimate each point's effect on model fairness.","marker":"[34]"},{"why":"Provides the upper confidence bound rule that balances exploration and exploitation in partition selection.","marker":"[5]"},{"why":"Provides the multi-armed bandit formulation the framework is built on.","marker":"[57]"},{"why":"Defines demographic parity and surveys fairness metrics used to measure the acquisition benefit.","marker":"[40]"},{"why":"Provides the ACS datasets used in the evaluation and the demographic parity target; introduced as fair ML benchmarks.","marker":"[22]"}],"fun_headline_variants":["Bandit learns which data to buy for fairer AI","Reinforcement learning picks data to boost fairness","Order data buys with bandits for fairer ML","RL ranks data points to fast-track fairness","Data-savvy bandit accelerates fairness gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the influence of a training point on fairness, computed with the current model, still predicts what happens when that point is actually added, and that a regressor can transfer that prediction to points in the pool; on top of that, DataSift's reward-sharing formula in Equation (1) is a heuristic rather than a derived fairness model.","fun_headline_variants_meta":{"raw":{"variants":["Bandit learns which data to buy for fairer AI","Reinforcement learning picks data to boost fairness","Order data buys with bandits for fairer ML","RL ranks data points to fast-track fairness","Data-savvy bandit accelerates fairness gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000855,"raw_usage":{"total_tokens":3733,"prompt_tokens":986,"completion_tokens":2747,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":2674}},"tokens_in":602,"tokens_out":2747,"duration_ms":19846,"temperature":1.0,"reasoning_tokens":2674,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:51:19.438118+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any of the six datasets and a random sample of unseen pool points; for each, retrain once with that single point added, record the actual change in demographic parity, and correlate those changes with the influences predicted by the ridge regressor. A near-zero or negative correlation would falsify the valuation layer, and then DataSift-Inf would reduce to random batch selection.","supporting_citations":[{"cited_title":"Selective data acquisition in the wild for model charging","cited_arxiv_id":null,"evidence_quote":"Defines MAB-based selective data acquisition for model accuracy; DataSift adapts this arm-selection idea to fairness and uses it as the AutoData baseline."},{"cited_title":"Understanding black-box predictions via influence functions","cited_arxiv_id":null,"evidence_quote":"Supplies the first-order influence functions that DataSift-Inf uses to estimate each point's effect on model fairness."},{"cited_title":"Slice tuner: A selective data acquisition framework for accurate and fair machine learning models","cited_arxiv_id":null,"evidence_quote":"Provides the multi-armed bandit formulation the framework is built on."},{"cited_title":"A Survey on Bias and Fairness in Machine Learning","cited_arxiv_id":null,"evidence_quote":"Defines demographic parity and surveys fairness metrics used to measure the acquisition benefit."}],"review_version":1}