{"id":"7daa73aa-e9d9-4a81-9d26-cd8794479b4c","arxiv_id":"2412.00316","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HiMoE adds a node-wise mixture-of-experts architecture and a ratio-based fairness metric to spatial-temporal forecasting, claiming large gains in both accuracy and cross-node consistency.","lead":"This paper proposes HiMoE, a forecasting model that uses a mixture of specialized expert networks and a new fairness metric to improve both accuracy and consistency across spatial nodes. It reports gains of 9.22% to 60.01% over baselines on four traffic, population, and air quality datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fairness claim rests entirely on the self-authored WMAE metric (Eq. 10), which is used both as training loss and evaluation metric. The paper never validates WMAE as a meaningful fairness measure; it is just a reweighted MAE that amplifies errors on low-volume nodes.","rationale":"The reader identified WMAE validity as the weakest assumption. I agree: the fairness claim is central and the metric is self-designed, used for training and evaluation, and lacks external validation. The missing code, seeds, and protocol mismatch are secondary concerns; the metric issue is more fundamental because it undermines the conceptual contribution. The proposed concrete test would settle whether the fairness gains are an artifact of the metric. If the test does not support WMAE, the verdict should remain CONDITIONAL, consistent with the reader; if it does, the paper's fairness contribution is strengthened. Thus the verdict is unchanged.","tokens_in":14717,"tokens_out":7022,"duration_ms":61901,"concrete_test":"On the PeMS04 dataset, re-evaluate every model in Table 2 (and Table 3) using two independent fairness definitions: (i) the standard deviation of per-node MAE, and (ii) the mean per-node MAPE (with a fixed small constant for zero denominators). Also train a simple reweighted-MAE baseline that minimizes Eq. 13 with the same WMAE loss but uses a standard architecture (e.g., Graph WaveNet). If HiMoE's SWMAE reductions do not carry over to the alternative fairness metrics, or if the reweighted Graph WaveNet matches HiMoE's SWMAE, then the fairness gains are metric artifacts and WMAE's validity is the decisive issue.","verdict_should_be":"UNCHANGED","load_bearing_attack":"WMAE (Eq. 10) is defined as mean(X)/mean(X_i) * (1/T') * sum_t |X_it - Xhat_it|. This is exactly a node-wise reweighted MAE with weight proportional to the inverse of the node's time-series mean. The claim that this 'eliminates the inherent bias of MAE' is unsupported: the transform changes the scale of the error but does not remove any bias; it simply emphasizes low-volume nodes. The paper uses WMAE for both training (Eq. 13) and evaluation (MWMAE/SWMAE), creating a circular argument. There is no external validation that reducing SWMAE corresponds to a meaningful fairness improvement, e.g., on nodes with rare events or severe underprediction. Moreover, the paper's assertion of being 'the first to identify node-level bias' is factually dubious given STMoE (Li et al., 2023) and FairSTG (Lin et al., 2024) already address performance heterogeneity. If WMAE is not a valid fairness metric, then STFairBench's contribution as a benchmark and the claimed fairness gains of HiMoE lose their basis, even if the architecture improves accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes HiMoE, a spatial-temporal forecasting framework combining a Heterogeneity-Informed Graph Convolutional Network (HiGCN) for trend heterogeneity and a Node-wise Mixture-of-Experts (NMoE) for cardinality heterogeneity. It also introduces STFairBench, a fairness benchmark with a new node-level metric WMAE and a training loss. Experiments on PeMS04, Beijing, Tongzhou, and KnowAir report that HiMoE outperforms seven baselines by 9.22%-60.01% across MAE, RMSE, MAPE, MWMAE, and SWMAE, with ablations, parameter sensitivity, efficiency, and case studies.","tokens_in":14996,"tokens_out":6159,"duration_ms":49932,"significance":"If the claims hold, the architecture is a plausible contribution to balancing accuracy and consistency in multi-node forecasting, and STFairBench could be a reusable evaluation framework. The efficiency analysis and the attempt to control for the training loss in Table 3 are useful steps. However, the headline comparison is confounded by training loss, and the fairness metric is used both for training and evaluation, so the scientific claims require careful validation before the work can be considered a reliable state-of-the-art result.","major_comments":[{"comment":"The state-of-the-art comparison is confounded by the training loss. HiMoE is trained with the STFairBench loss (Eq. 13) while each baseline uses its original loss. Since the loss directly controls the trade-off between accuracy and consistency, the improvements in MWMAE and SWMAE may be attributed to the loss function rather than the HiMoE architecture. Section 4.4 (Table 3) retrains baselines with the fairness loss, but only on PeMS04, so the abstract's claim of 'outperforming the best baseline by at least 9.22% across all evaluation metrics' is not supported by a like-for-like comparison on the other three datasets. Please provide controlled comparisons across all datasets, or restrict the claims appropriately.","section":"Section 4.2, Table 2"},{"comment":"The fairness metric WMAE is authored in this paper and is used both as the training objective (Eq. 13) and as the evaluation metric (MWMAE/SWMAE). This creates a circularity that makes the reported fairness gains partly self-confirming. The paper does not validate that WMAE is a meaningful fairness metric; it is a reweighted MAE with weights inverse to the node's mean, and the claim that MAE is 'inherently biased' is not formally justified. Please provide an external validation, for example by showing that SWMAE correlates with errors on low-volume nodes or by comparing against another fairness definition from the literature.","section":"Section 3.3, Eqs. (10)-(13)"},{"comment":"The experiments report single runs with no error bars, seeds, or significance tests. Consequently, it is unclear whether the large margins over baselines are within run-to-run variance. The paper states that implementation details are in a code repository, but no repository link is provided. Please add code and detailed experimental settings (data splits, seeds, hyperparameters) and report standard deviations over at least three runs.","section":"Section 4.2-4.6"},{"comment":"The paper claims that 'HiMoE is the first method to identify node-level bias in spatial-temporal predictions arising from existing training strategies.' This is contradicted by the paper's own related work (Section 5.1), which cites STMoE (Li et al., 2023) as mitigating performance bias across road segments and FairSTG (Lin et al., 2024) as targeting performance heterogeneity. Please revise the novelty claim to a precise statement and describe the specific differences from these methods.","section":"Introduction and Section 6"}],"minor_comments":[{"comment":"The definition of WMAE in Eq. (10) states that 'both X_i and X represent the entire time series,' but the next sentence says the training loss excludes test data. Please clarify which data splits are used to compute the scaling factors for the training loss versus the evaluation metric.","section":"Section 3.3"},{"comment":"The text says 'applying sparsity over the expert dimension' after Eq. (7), but the softmax does not zero out any expert probabilities. If sparsity is achieved by top-k selection or by the sigmoid output gating in Eq. (9), please state the explicit sparsification mechanism.","section":"Section 3.2.2"},{"comment":"Figure 5 and Figure 6 are informative, but the sensitivity analysis would be strengthened by reporting numerical values or error bars, especially for the claim that 'increasing k consistently improves prediction performance'.","section":"Section 4.5"},{"comment":"The statement 'HiMoE is the first to adopt a fine-grained expert mixture framework' is too strong without a formal definition of 'fine-grained' and given the existing MoE-based spatial-temporal models cited in the paper.","section":"Section 4.6"},{"comment":"The paper references a code repository for implementation details, but no URL is given in the text or footnotes.","section":"Section 4.1.4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core technical idea is interesting, and the efficiency and ablation studies are useful. However, the uncontrolled comparison and the circularity of the fairness metric need to be resolved before the claims can be trusted. I believe a major revision with controlled experiments and metric validation could bring the paper to acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: HiMoE is a solid, well-motivated architecture for reducing node-level disparities in spatial-temporal forecasting. The combination of heterogeneity-informed graph convolution, node-wise MoE routing, and a weighted-MAE training loss is a real, if incremental, contribution. But the paper's fairness claims rest on a self-authored metric—WMAE—that is essentially a node-wise reweighted MAE, and the paper never shows that minimizing its spread corresponds to fairness in any externally meaningful sense. The headline comparison in Table 2 is also unfair: HiMoE trains with its own loss while baselines use their original losses. Table 3, which retrains baselines with the proposed loss, is the right experiment and should be the main table.\n\nWhat's genuinely good: the architecture is coherent and the ablations show each piece matters. The efficiency table is honest about the parameter count and shows training is still fast. The case study on KnowAir gives a concrete visual of the fairness problem. The authors also include a caveat that results don't fully reflect each method's effectiveness. I'd credit the empirical discipline.\n\nThe soft spots are real but not fatal. WMAE as defined in Eq. 10 is mean(X)/mean(X_i) times the node MAE—so it's a per-node relative error scaled back to data magnitude. Calling it 'unbiased' is a stretch; it's a reweighting that emphasizes low-volume nodes. Using it for both training and evaluation is circular to some degree, though not uniquely damning—many works train and evaluate on the same metric. The accuracy gains (MAE/RMSE/MAPE) are independent and large, so the architecture clearly does something. The bigger problems are the missing code, seeds, and error bars, and the 'first to identify node-level bias' claim, which ignores STMoE and FairSTG already addressing performance heterogeneity. The related work section cites them, so the overclaim is in the framing, not the bibliography.\n\nWho's this for? Researchers working on fairness in urban computing or ST forecasting. It's a useful benchmark idea and a plausible model, but I wouldn't stake a project on it until code and error bars appear. I'd give it peer review, not a desk rejection, but with a clear request for major revision: release code, report multiple seeds, make Table 3 the central comparison, and either validate WMAE against an external fairness criterion or soften the fairness language.","headline":"A credible architecture for fair ST forecasting whose fairness metric is a reweighted MAE, with an unfair head-to-head in the main table—but the accuracy gains are independent and the paper deserves peer review.","tokens_in":15541,"tokens_out":2670,"would_cite":false,"duration_ms":25218,"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 existing forecasting losses are biased against low-activity nodes, and presents HiMoE plus a ratio-based fairness benchmark that beats the best baseline by at least 9.22% on every metric across four urban datasets.","keywords":["spatial-temporal forecasting","fairness","mixture-of-experts","graph convolutional network","node-level bias","cardinality heterogeneity","trend heterogeneity","STFairBench"],"falsifier":"Build a synthetic or real dataset where all nodes have identical means, train HiMoE with the STFairBench loss, and check whether SWMAE still drops substantially relative to MAE training; if it does, the cardinality-normalization story is not the source of the gain. Alternatively, replace WMAE with MAPE plus a small epsilon in the same loss; if the SWMAE improvements disappear, the specific WMAE ratio is load-bearing, and if they remain, the fairness contribution of the metric is weaker than claimed.","tokens_in":14466,"feed_emoji":"🚦","tokens_out":7952,"duration_ms":69979,"temperature":0.7,"pith_summary":"The paper sets out to establish that fairness in spatial-temporal forecasting—consistent accuracy across nodes—can be improved without sacrificing overall accuracy. It argues that standard averaged losses are biased because nodes with smaller average magnitudes show larger relative errors under the same absolute error, so low-activity sensors and regions are systematically underserved. To fix this, it proposes HiMoE, which routes nodes to specialists by cardinality and models trend-based correlations between nodes, and STFairBench, a benchmark whose weighted-MAE metric and loss penalize node-level error spread. The central empirical claim is that on four urban datasets covering traffic flow, speed, population, and air quality, HiMoE beats every compared baseline on every metric—including the two fairness metrics—by at least 9.22%.","feed_headline":"Forecasting model beats best baseline by 9% to 60% on four datasets","feed_subtitle":"Node-wise experts plus a ratio-based fairness metric keep predictions accurate and consistent across sensors.","key_machinery":"The load-bearing object is the weighted MAE of Eq. 10, $\\mathrm{WMAE}_i = \\frac{\\mathrm{mean}(X)}{\\mathrm{mean}(X_i)} \\cdot \\frac{1}{T'} \\sum_j |X_i^j - \\hat{X}_i^j|$, which rescales each node's absolute error by its own mean value so that low-cardinality and high-cardinality nodes are compared on the same relative footing; MWMAE is its mean across nodes and SWMAE its standard deviation, and the training loss $L = \\mathrm{MWMAE} + \\alpha\\,\\mathrm{SWMAE}$ is what presses the model toward consistency. The architecture that carries the accuracy side is HiGCN, which fuses a Gaussian-kernel static adjacency matrix with a dynamic trend-similarity adjacency matrix and applies a tanh edge gate before spectral graph convolution, and NMoE, which learns cardinality representations, routes nodes to experts through a temperature-controlled softmax over expert centers, and prunes inter-node edges inside each expert so that nodes of very different magnitudes are not forced to depend on each other.","core_discovery":"The paper's central claim is that node-level heterogeneity has two distinct forms and both must be addressed jointly to get accurate and fair spatial-temporal forecasts. Trend heterogeneity—different daily patterns across nodes—requires graph convolutions that fuse static distance graphs with dynamic trend-similarity graphs. Cardinality heterogeneity—different average magnitudes across nodes—requires assigning nodes to experts by magnitude so that low-volume nodes are not modeled together with high-volume nodes. The paper further claims that MAE and RMSE are unfair for this purpose, and that its WMAE metric, which normalizes each node's error by its own mean, gives an unbiased node-level fairness signal. It reports state-of-the-art results on the PeMS04, Beijing, Tongzhou, and KnowAir datasets, with improvements from 9.22% to 60.01% across MAE, RMSE, MAPE, MWMAE, and SWMAE, and it claims to be the first method to identify and correct node-level bias in spatial-temporal prediction.","pith_inferences":["A natural extension is to test whether minimizing SWMAE also reduces worst-case error on low-activity nodes or merely shifts error toward medium-activity nodes, since the metric alone does not distinguish those outcomes.","The gating is driven by cardinality embeddings, which suggests a testable transfer experiment: train expert centers on one city and route a second city's sensors without retraining, checking whether the learned cardinality ordering transfers.","The WMAE construction is close to a relative error, so the benchmark could plausibly generalize beyond traffic and air quality to other spatiotemporal tasks with strong cardinality heterogeneity, such as energy load or mobility demand forecasting."],"forward_implications":["If the WMAE-based benchmark is accepted, future fairness results in spatial-temporal forecasting should report MWMAE and SWMAE alongside MAE, RMSE, and MAPE.","If the fairness loss works as claimed, training with MWMAE plus alpha times SWMAE should reduce cross-node error spread without sacrificing average accuracy.","If the empirical results hold, fine-grained mixture-of-experts with many small experts is a viable scaling path for spatial-temporal forecasting, not just for large language models.","If the ablations reflect the true mechanism, both the trend-based dynamic graph fusion and the node-wise input gating are necessary for the reported gains."],"supporting_citations":[{"why":"The existing fairness-aware mixture-of-experts baseline whose reported results HiMoE is compared against and must beat.","marker":"[24]"},{"why":"The prior fairness method that reweights nodes using aggregate metrics, the approach the paper argues is biased.","marker":"[25]"},{"why":"The fine-grained sparse mixture-of-experts design that NMoE adapts to spatial-temporal nodes.","marker":"[26]"},{"why":"The foundational sparsely-gated mixture-of-experts mechanism that the node-wise gating extends.","marker":"[31]"},{"why":"A widely used adaptive-graph traffic forecasting baseline used in the comparison.","marker":"[38]"},{"why":"An attention-based spatial-temporal graph baseline that is the strongest competitor on PeMS04.","marker":"[11]"},{"why":"A graph-based multivariate time-series baseline used to compare graph learning.","marker":"[37]"},{"why":"A time-enhanced mixture-of-experts forecasting baseline that NMoE is contrasted with.","marker":"[23]"}],"fun_headline_variants":["Heterogeneity-aware experts boost fair forecasting by up to 60%","Per-node expert routing yields fairer forecasts with up to 60% gains","Mixture-of-experts model fixes node bias for accurate fair ST forecasting","HiMoE: Heterogeneity-aware mixture-of-experts for fair forecasts","Node-wise experts and a new metric deliver fairer spatial-temporal predictions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that WMAE, which divides each node's absolute error by its own mean value, is an unbiased measure of node-level prediction quality, so that reducing its cross-node spread is genuine fairness rather than just reshaping errors to fit a chosen ratio.","fun_headline_variants_meta":{"raw":{"variants":["Heterogeneity-aware experts boost fair forecasting by up to 60%","Per-node expert routing yields fairer forecasts with up to 60% gains","Mixture-of-experts model fixes node bias for accurate fair ST forecasting","HiMoE: Heterogeneity-aware mixture-of-experts for fair forecasts","Node-wise experts and a new metric deliver fairer spatial-temporal predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000613,"raw_usage":{"total_tokens":2847,"prompt_tokens":942,"completion_tokens":1905,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":1809}},"tokens_in":558,"tokens_out":1905,"duration_ms":13593,"temperature":1.0,"reasoning_tokens":1809,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:31:21.419840+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a synthetic or real dataset where all nodes have identical means, train HiMoE with the STFairBench loss, and check whether SWMAE still drops substantially relative to MAE training; if it does, the cardinality-normalization story is not the source of the gain. Alternatively, replace WMAE with MAPE plus a small epsilon in the same loss; if the SWMAE improvements disappear, the specific WMAE ratio is load-bearing, and if they remain, the fairness contribution of the metric is weaker than claimed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The foundational sparsely-gated mixture-of-experts mechanism that the node-wise gating extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A time-enhanced mixture-of-experts forecasting baseline that NMoE is contrasted with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The prior fairness method that reweights nodes using aggregate metrics, the approach the paper argues is biased."}],"review_version":1}