{"id":"77f14952-e8a3-47c7-9f53-43b03ac4fadb","arxiv_id":"2412.16214","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"FairTP adds dynamic fairness losses and state-guided sampling to traffic prediction, but its evaluation is weakened by uncontrolled comparisons and inconsistent reporting.","lead":"The paper introduces FairTP, a framework that adds fairness constraints and a sampling strategy to traffic prediction models, aiming to keep prediction errors balanced across city regions and sensors over time. It reports improved fairness metrics on two real-world datasets, but the experimental comparisons are confounded by using different amounts of training data and by defining fairness metrics that only its own framework can compute.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prolonged-fairness claim rests on a self-referential metric: SDF is computed from FairTP's own discriminator states, with no baseline values or external validation.","rationale":"The reader's weakest_assumption identifies exactly the state-identification/SDF circularity, and I agree that it is load-bearing. The central contribution is prolonged fairness, and the only metric for it is generated by the framework itself; without baseline values or external validation, the abstract's claim cannot be checked. I additionally note the Table 1/Table 4 inconsistencies, which strengthen the REJECT verdict, but the primary reason for rejection remains the unverifiable SDF. A matched-sampling baseline plus ground-truth SDF computation would resolve this: if FairTP's SDF gains vanish, the paper should be rejected; if they survive, the concern is addressed. The verdict should therefore remain UNCHANGED from the reader's REJECT.","tokens_in":14083,"tokens_out":4123,"duration_ms":38574,"concrete_test":"Using the released code, recompute SDF on the held-out test set with ground-truth states defined by comparing each sensor's MAPE to the baseline threshold, and compute this ground-truth SDF for FairTP and for the unmodified baseline trained on the same stratified 200-sensor sample. If FairTP does not beat the matched baseline on this externally computed SDF, the prolonged-fairness improvement is a discriminator artifact rather than a verified fairness gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FairTP significantly improves prolonged fairness is supported only by SDF numbers in Table 1 and the ablation, yet SDF is not an observable quantity. In Eqs. (3)-(5), d_{t_k}^{v_i} is the discriminator's predicted state, and the labels used to train the discriminator are produced by thresholding MAPE against the original baseline's per-round MAPE (Appendix C). Thus SDF measures agreement of a discriminator with labels derived from a baseline, not a directly validated fairness property of the predictions. The paper explicitly says baselines 'cannot produce SDF outputs', so there is no baseline value to improve upon. Moreover, LSDF has a degenerate minimizer: a discriminator that predicts a constant state for all sensors makes every D_{T_d} equal and LSDF = 0, while the sampling module then receives uniform probabilities; the paper does not analyze or regularize this collapse. The ablation in Table 3 shows SDF swings from 0.06 to 188 when LSDF is removed, which only shows the loss changes the discriminator, not that states track real sacrifice/benefit. The numerical inconsistencies between Table 1 and Table 4 (e.g., FairTP-DCRNN HK MAE +2.46% in Table 1 but reported as -1.20% in Table 4) further undermine the accuracy side of the claim, but the core unverifiability is the self-referential SDF.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FairTP, a framework that augments traffic forecasting models with a region-based static fairness (RSF) loss, a sensor-based dynamic fairness (SDF) loss, a state identification module that labels sensors as \"sacrifice\" or \"benefit\" via a discriminator, and a state-guided balanced sampling strategy. The authors claim that on the HK and SD datasets FairTP significantly improves fairness while minimizing accuracy degradation, and they report comparisons across seven traffic prediction backbones, two fairness mitigation baselines, ablations, and parameter analyses.","tokens_in":14316,"tokens_out":3673,"duration_ms":34934,"significance":"The problem of fairness over time in traffic prediction is timely and practically relevant, and the paper makes a concrete attempt by releasing code, using two real-world datasets, and testing seven backbone models. If the claimed effects were reliably established, FairTP would be a useful plug-in for equitable traffic forecasting. However, the current manuscript does not verify the central claim: the SDF metric is computed from the model's own discriminator, no baseline SDF values exist, and the main comparison is confounded by an asymmetry in the amount of training data. As presented, the evidence does not support the conclusion that FairTP improves prolonged fairness without accuracy loss.","major_comments":[{"comment":"The comparison is uncontrolled: Appendix D.3 states that \"for all baselines without FairTP, we use all sensors for training,\" while FairTP uses Nsam=200 sampled sensors. Consequently, every FairTP-versus-baseline comparison in Tables 1, 2, and 4 simultaneously changes the amount of training data and the model components, so the reported MAE and RSF differences cannot be attributed to the proposed fairness mechanism. A controlled comparison using the same sensor subset for baselines, or a FairTP variant trained on all sensors, is required before any accuracy or fairness claim can be evaluated.","section":"Comparison With Baselines and Appendix D.3"},{"comment":"SDF as defined in Equations (3)-(4) is not an observable property of the predictions; it is computed from the discriminator outputs d_tk_vi, and the state labels are generated by thresholding per-sensor MAPE against the original model's per-round MAPE (Appendix C). Since the same module is optimized by LSDF and used to measure the outcome, the reported SDF improvements in Table 1 and Table 3 are self-referential. The paper explicitly states that baselines \"cannot produce SDF outputs,\" so no baseline SDF values exist to substantiate the prolonged-fairness improvement; an external validation based directly on prediction-error dynamics over time is needed.","section":"Sensor-based Dynamic Fairness and State Identification Module"},{"comment":"The SDF loss has a degenerate minimizer: if the discriminator outputs a constant state for all sensors, every pairwise difference D_Td[vi] - D_Td[vj] is zero, making LSDF equal to zero, and the sigmoid in Equation (5) then produces uniform sampling probabilities. The paper does not analyze or regularize this collapse. The ablation in Table 3, where SDF swings from 0.06 to 188 when LSDF is removed, only shows that the loss changes discriminator outputs; it does not show that the discriminator's states track actual sacrifice/benefit of sensors.","section":"Equation (4) and LSDF"},{"comment":"The reported accuracy changes are numerically inconsistent. For example, the text discussing Table 1 reports FairTP-DCRNN on HK as having MAE +2.46%, while Table 4 reports -1.20%; FairTP-AGCRN HK MAE is +4.91% in Table 4, whereas the raw MAE values in Table 1 (1.939 vs. 1.957) correspond to approximately -0.9%. These discrepancies undermine the accuracy side of the central claim and must be corrected, with a precise statement of how the percentage changes are computed.","section":"Table 1 versus Table 4"}],"minor_comments":[{"comment":"Equation (1) defines M[yhat_t^rp] as the mean absolute percentage error for region rp at time t, but MAPE cannot be computed from a single time point without specifying an error window or aggregation procedure; please clarify.","section":"Equation (1)"},{"comment":"The text says that when LSDF is removed \"the SDF decreases remarkably,\" but Table 3 shows that removing LSDF (noD) makes SDF much larger (worse); the wording appears to be backwards and should be corrected.","section":"Ablation Study discussion"},{"comment":"The SDF column in Table 2 contains only dashes and is never discussed; if SDF is not computable for FairST and SA-Net, this should be stated in the caption or text to avoid confusion.","section":"Table 2"},{"comment":"The notation in Equation (9) is confusing: d is described as the predicted state and Y as the corresponding label, but the cross-entropy form mixes predicted and label roles; please use consistent notation such as y_pred and y_label.","section":"Equation (9)"},{"comment":"Table 3 lists \"FairTP-G2STGNN,\" which appears to be a typo for \"FairTP-D2STGNN\"; please correct it.","section":"Table 3"}],"recommendation":"reject","confidential_remarks":"The central evaluation issues are not local fixes: the prolonged-fairness metric is defined through the proposed model's own components, and the main comparison changes data quantity along with method. Even with corrected baselines, the SDF evidence would remain unverifiable without an external definition of prolonged fairness. I do not see a path within the current manuscript's scope that would establish the abstract's claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it expecting a useful extension of fairness to the temporal dimension, and that part is genuinely there. The problem statement is valid: sensors are unevenly deployed, and static fairness can erode as traffic conditions change. Defining two fairness notions (RSF for regional disparity, SDF for sensor-state dispersion over a window) and building a sampling module on top is a reasonable research direction. Running the framework over seven underlying traffic models and two fairness baselines shows effort, and they posted code.\n\nBut the evaluation as reported does not support the abstract's claim. Two issues are load-bearing.\n\nFirst, the comparison is not controlled. FairTP trains on Nsam=200 sampled sensors, while every baseline trains on the full sensor set (stated in Appendix D.3). That changes the training-data distribution and model capacity, so the accuracy differences between FairTP and baselines cannot be attributed to the fairness mechanism. Either compare all methods on the same sensor budget or at least provide a baseline trained on 200 sensors.\n\nSecond, the prolonged-fairness metric SDF is computed only for FairTP and is defined through the framework itself. State labels are produced by thresholding each sensor's MAPE against a threshold fixed from the unmodified baseline's per-round MAPE (Appendix C). The discriminator then predicts those labels, and SDF is derived from the predicted states. So SDF measures how well the internal discriminator agrees with a fixed baseline-derived rule, not an independently validated fairness property. Baselines are said to be unable to produce SDF, so there are no baseline SDF values to improve on. That makes the main 'prolonged fairness' claim unverifiable.\n\nThere are also two more concrete problems. LSDF has a trivial zero solution: a discriminator that predicts the same state for every sensor makes all D_Td equal and LSDF = 0. The paper doesn't analyze or regularize this collapse. The ablation's SDF swings (0.06 to 188 when LSDF is removed) are consistent with the loss reshaping the discriminator's outputs rather than tracking real sacrifice/benefit. And the numbers are internally inconsistent: Table 1 and Table 4 contradict each other, e.g. FairTP-AGCRN on HK shows MAE 1.939 vs 1.957 (an improvement) in Table 1, but Table 4 reports a +4.91% degradation.\n\nThe conceptual direction is worth pursuing, but the current evidence doesn't justify the claims. I'd give this a reject now, with a clear path to a better paper: controlled comparisons, an external SDF definition computable for baselines, corrected numbers, and an analysis of the degenerate discriminator. It is a good paper to put in front of a reading group as a case study in self-referential metrics. For peer review, I'd still send it out—the idea is novel enough and the problem important enough that referees should see it, and the flaws are fixable in revision.","headline":"Prolonged fairness is a real gap, but the new SDF metric is self-referential and the evaluation is confounded by uneven sensor budgets; the current evidence doesn't support the paper's headline claims.","tokens_in":14907,"tokens_out":6047,"would_cite":false,"duration_ms":50457,"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":"FairTP aims to make traffic predictions fair across regions and over time by defining sensor states and rebalancing training data.","keywords":["traffic prediction","algorithmic fairness","prolonged fairness","region-based static fairness","sensor-based dynamic fairness","state-guided balanced sampling","state identification","spatio-temporal graph neural network"],"falsifier":"Recompute SDF on the test set using ground-truth state labels (each sensor's actual error compared with the threshold) instead of the discriminator's predictions, and run FairTP with Td = 24 instead of 3; if the reported SDF improvements disappear or reverse in either check, the prolonged-fairness claim is not supported.","tokens_in":13819,"feed_emoji":"⚖️","tokens_out":8798,"duration_ms":70901,"temperature":0.7,"pith_summary":"FairTP is a plug-in framework for traffic prediction that aims to make forecasts fair across city regions and over time. Existing fairness methods for traffic typically constrain predictions only at single time points, so fairness can break as traffic conditions change. FairTP defines two fairness metrics: region-based static fairness (RSF), the average absolute difference in prediction-error rates between pairs of regions at a time point, and sensor-based dynamic fairness (SDF), the absolute difference between sensors' accumulated 'sacrifice' and 'benefit' states over a short period. A state identification module labels each sensor as 'sacrifice' or 'benefit' and a state-guided sampling module gives more training weight to sacrificing sensors. Experiments on two real-world traffic datasets with seven base models report that FairTP improves both fairness metrics while keeping mean absolute error roughly unchanged or better.","feed_headline":"FairTP keeps traffic predictions fair over time","feed_subtitle":"It adds two fairness measures and balanced sampling to any traffic model, with little accuracy loss.","key_machinery":"The mechanism that carries the argument is the state identification module plus a state-guided balanced sampling loop. A state marker converts each sensor's current MAPE into a binary label by comparison with a threshold recorded from the original model during its own training; a discriminator then learns to predict this label from the spatio-temporal model's hidden representation H_st, so states can be assigned at test time. Every Td batches, the framework accumulates the predicted states into an overall state per sensor, converts the sum into a sampling probability through a sigmoid, combines it with region-level balancing probabilities, and greedily selects the Nsam sensors with the lowest probabilities for the next training period. The fairness objectives are L_RSF, the average absolute regional MAPE difference at each time point, and L_SDF, the average absolute difference of accumulated states over the window, added to the prediction loss as λ1 L_RSF + λ2 L_SDF.","core_discovery":"The paper's central claim is that prolonged fairness in traffic prediction can be defined and optimized without sensitive attributes. Each sensor is placed in one of two states by comparing its mean absolute percentage error (MAPE) with a threshold taken from the unmodified base model: 'benefit' (error below threshold) or 'sacrifice' (error above it). Fairness over a period is achieved when the sums of these states across sensors are similar, which is the sensor-based dynamic fairness (SDF) measure; region-based static fairness (RSF) handles the per-time-point regional gap. The FairTP framework realizes this by co-training a discriminator that infers sensor states from the base model's hidden representations, and by using those states to rebalance training samples every Td batches. On the HK and SD datasets, the paper reports that FairTP attached to DCRNN, AGCRN, GWNET, ASTGCN, DSTAGNN, DGCRN, and D2STGNN lowers RSF and SDF while keeping mean absolute error nearly unchanged or improved in most configurations.","pith_inferences":["Going beyond the paper, the threshold used to label 'sacrifice' and 'benefit' is calibrated once on the unmodified model; as FairTP's sampling changes the training distribution, that threshold can become stale, so a natural extension is periodic re-calibration and a check of whether SDF gains persist.","Going beyond the paper, the definition of 'prolonged' is tied to the window Td=3 chosen in experiments; testing the framework on daily or weekly windows would show whether the fairness guarantee actually extends to the time scales that matter for transport policy.","Going beyond the paper, the same state-identification-plus-rebalancing recipe could be applied to other spatio-temporal forecasting tasks, such as energy or ride-hailing demand, where zones alternate between well-served and under-served error regimes."],"forward_implications":["A traffic agency can attach FairTP to an existing spatio-temporal predictor without redesigning the model, because the ST module is replaceable.","The fairness metrics need no sensitive attributes such as race or gender, so they apply directly to traffic sensor data.","Because sampling is rebalanced every Td batches toward sensors in 'sacrifice', predictions in under-served regions can improve while dense-sensor regions do not lose much accuracy, per the regional performance analysis.","The composite loss gives a tunable trade-off between accuracy, static regional fairness, and prolonged sensor fairness via λ1 and λ2.","Compared with the static fairness methods FairST and SA-Net on AGCRN and D2STGNN backbones, the paper reports lower regional disparity with comparable or lower MAE."],"supporting_citations":[{"why":"DCRNN is one of the seven spatio-temporal base models and its per-round MAPE provides the threshold for state labels.","marker":"Li et al. 2018"},{"why":"AGCRN is the backbone used in the main comparisons, ablations, and parameter analyses.","marker":"Bai et al. 2020"},{"why":"GWNET is another base model demonstrating FairTP's integration with a temporal-convolution model.","marker":"Wu et al. 2019"},{"why":"FairST is the static fairness baseline that FairTP must outperform.","marker":"Yan and Howe 2020"},{"why":"SA-Net is the second fairness baseline, contributing its fairness-regularization integration approach.","marker":"Zheng et al. 2023a"},{"why":"D2STGNN is a decoupled spatio-temporal backbone used in the comparison and ablation tables.","marker":"Shao et al. 2022"}],"fun_headline_variants":["FairTP keeps traffic forecasts fair over the long haul","Prolonged fairness for traffic predictions, not just at one time","Traffic prediction fairness that endures across time and regions","Two fairness measures keep traffic predictions fair over time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a sensor's 'sacrifice' or 'benefit' state, defined by comparing its current prediction error rate with a threshold from the original model, can be reliably predicted at test time from the hidden representation, and that summing these predicted states over a three-batch window captures prolonged fairness.","fun_headline_variants_meta":{"raw":{"variants":["FairTP keeps traffic forecasts fair over the long haul","Prolonged fairness for traffic predictions, not just at one time","Traffic prediction fairness that endures across time and regions","Two fairness measures keep traffic predictions fair over time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000925,"raw_usage":{"total_tokens":4010,"prompt_tokens":1037,"completion_tokens":2973,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":2906}},"tokens_in":653,"tokens_out":2973,"duration_ms":19650,"temperature":1.0,"reasoning_tokens":2906,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:08:57.693698+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute SDF on the test set using ground-truth state labels (each sensor's actual error compared with the threshold) instead of the discriminator's predictions, and run FairTP with Td = 24 instead of 3; if the reported SDF improvements disappear or reverse in either check, the prolonged-fairness claim is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FairST is the static fairness baseline that FairTP must outperform."}],"review_version":1}