{"id":"cc47d584-eb39-40ad-95b2-f99b3aa3921d","arxiv_id":"2501.06255","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A label-decomposition training scheme with per-component supervision and recombination, combined with random subgraph sampling, improves long-term large-scale wireless traffic forecasting.","lead":"This paper proposes PSLD, a training method that decomposes the future traffic matrix into mean, variance, and residual components, supervises a neural network on each component separately, and then recombines them for long-term wireless traffic forecasting. It also introduces a random subgraph sampling strategy for city-scale base station graphs, with claimed 2-11% error reductions on three datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I does not support the claimed 2%/4%/11% averages: on C2TM the best-baseline gain is ~0.06%, and the 11% on CBS only appears versus PatchTST, not versus stronger baselines.","rationale":"The reader's weakest assumption focused on whether label decomposition yields easier-to-learn components, which is a mechanism-level concern. My concern is more direct: the paper's own Table I does not reproduce the headline percentage improvements. The central claim of the paper is empirical superiority over SOTA, and that claim is falsified as stated by the numbers in its main table. This is more load-bearing than the 'progressive supervision' naming issue, because even if components are easier to learn, the reported gains must be real and correctly aggregated. I credit the authors for releasing WTFlib and code, which makes the test concrete and feasible. The verification should first check whether the aggregate percentages are computed against the best baseline per dataset or against a weaker baseline; if the former, C2TM collapses to a 0.06% improvement and CBS to 4.8%, and multi-seed error bars would probably show non-significance. I also flag that Theorem 1 is unproven in the available text and that Algorithm 1's pseudocode does not implement subgraph splitting; these are secondary but should be corrected. Since the reader already assigned CONDITIONAL, my finding reinforces that verdict without moving it; the paper should be accepted only after the empirical claims are recomputed and reported with uncertainty estimates.","tokens_in":18509,"tokens_out":7608,"duration_ms":72283,"concrete_test":"Recompute dataset-level averages from Table I: for each dataset, take the mean of PSLD-MVD MSE over the listed prediction lengths and the mean of each baseline; record the best baseline and the relative improvement. Then run PSLD, FreTS, TimeMachine, and PatchTST with 5 random seeds on C2TM, Milano, and CBS and report mean±std. Check whether the C2TM gain (currently ~0.06%) and the CBS gain over TimeMachine (~4.8%) are statistically significant. If the 2%/11% figures come from a different comparison than best-overall-baseline, the paper must state the exact baseline and aggregation used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (abstract; Section IV.D) is that PSLD improves over SOTA by average 2%, 4%, and 11% on C2TM, Milano, and CBS. Reproducing the aggregate from Table I does not support these figures. Averaging MSE across the reported prediction lengths, PSLD-MVD is 9.266 on C2TM versus 9.272 for the best baseline FreTS, a 0.06% improvement, not 2%. On Milano, PSLD-MVD is 0.875 versus 0.912 for TimeMachine (4.0%, matching the claim). On CBS, PSLD-MVD is 1.797 versus 1.887 for TimeMachine (4.8%); the 11% figure is only obtained by comparing to PatchTST (2.026), a weaker baseline on this dataset. Since no error bars or seeds are reported, a gap of 0.001 MSE on C2TM length-5 (both 9.161) is not evidence of significant improvement. The claim of 'significantly outperforming' is therefore not a well-defined, reproducible comparison. This is load-bearing because the paper's contribution is precisely this empirical superiority; if the aggregates are redefined or the gaps are within noise, the headline claim collapses. The table also contains anomalies, e.g., FourierGNN on CBS has MSE 1.193 at length-72 versus 2.601 at length-24, and C2TM PSLD-MVD MAE spikes to 0.317 at length-7, suggesting unreviewed numbers. A separate issue is that Theorem 1's proof is cited to a missing Appendix II, and Algorithm 1's pseudocode does not show subgraph partitioning, but the empirical discrepancy alone is sufficient to question the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PSLD (Progressive Supervision via Label Decomposition) for long-term, large-scale wireless traffic forecasting. It introduces Random Subgraph Sampling (RSS) to reduce the scale of graph-structured traffic data, decomposes the target label into mean/variance/residual or trend/seasonal/residual components, trains separate learner/predictor branches on each component, and recombines their outputs through a combinatory module. The authors report experiments on three wireless traffic datasets (C2TM, Milano, CBS) and several LTSF benchmarks, claiming average performance improvements of 2%, 4%, and 11% over existing methods, together with faster training and inference. They also release WTFlib, an open-source library for wireless traffic forecasting.","tokens_in":18944,"tokens_out":5084,"duration_ms":46650,"significance":"If the claims hold, the paper would offer a simple and computationally cheap way to improve long-term forecasting on non-stationary, large-scale traffic data, and RSS would provide a practical route to training on very large graphs. The paper has clear strengths: the decomposition/recombination equations are exact when the combinatory layer is an identity mapping; the authors release code and data links; and the efficiency comparison (Table IV) is a useful contribution. However, the central empirical claim of superiority over SOTA is not reproducible from the reported numbers, the theoretical statement for RSS lacks its proof, and the word 'progressive' is not backed by any progressive training mechanism. The significance of the contribution therefore cannot be assessed until these load-bearing issues are resolved.","major_comments":[{"comment":"The claim of average 2%, 4%, and 11% improvement over SOTA is not reproducible from Table I. Averaging MSE over the reported horizons on C2TM gives 9.266 for PSLD-MVD versus 9.272 for the best baseline FreTS, a 0.06% gain; on Milano the PSLD-MVD average is 0.875 versus 0.912 for TimeMachine (4.0%); on CBS the PSLD-MVD average is 1.797 versus 1.887 for TimeMachine (4.8%), and the 11% figure is obtained only by comparing with PatchTST (2.026), which is not the strongest baseline on that dataset. Since no error bars or random seeds are reported, a 0.001 MSE difference at individual horizons (e.g., C2TM length-5, both 9.161) is not evidence of significant improvement. The headline claim therefore needs to be recomputed against the best baseline per dataset and accompanied by variance estimates.","section":"Abstract; Section IV.D, Table I"},{"comment":"Theorem 1 is the sole theoretical justification for RSS, but its proof is cited to 'Appendix II,' which is not present in the manuscript; the claim that sampled subgraphs are an unbiased estimator of true aggregate features is therefore unverifiable. Algorithm 1 also does not implement what the text describes: line 1 overwrites Nsub with Nnode/Nsub, and lines 7-8 index edge matrices by shuffled node indices without specifying how subgraphs are partitioned or selected, making the pseudocode incomplete.","section":"Section II, Theorem 1 and Algorithm 1"},{"comment":"The method is called 'progressive supervision,' but no progressive or curriculum mechanism is present. The component losses L_theta, L_phi, L_psi in Eq. (7) are summed with the combinatory loss in Eq. (9) from the first epoch with a fixed weight lambda; there is no staged training, no ordering of components by difficulty, and no schedule. The word 'progressively' in the description ('learned progressively at shallow layers') is not matched by any algorithmic step. To support the name and the claimed benefit of learning easy components first, the authors should compare against a non-progressive multi-task baseline that trains the same branches simultaneously with the same total loss.","section":"Section III, Eq. (9), Fig. 3"},{"comment":"The reported numbers contain anomalies that suggest unreviewed errors. For FourierGNN on CBS, MSE drops from 2.601 at length-24 to 1.193 at length-72, while most models degrade with horizon. For PSLD-MVD on C2TM, the MAE at length-7 is 0.317 while all other MAEs are about 0.17, yet the MSE at the same length is 9.431, consistent with the neighboring values. These entries need to be verified and corrected; if they are typographical errors, the table must be regenerated.","section":"Section IV.D, Table I (FourierGNN CBS row; PSLD-MVD C2TM MAE row)"}],"minor_comments":[{"comment":"The title reads 'An Long-Term' and should be 'A Long-Term'; Section IV.D contains 'including including' in the sentence listing FreTS, TimeMachine, and FourierGNN.","section":"Title; Section IV.D"},{"comment":"The symbols M and S in Eq. (12) denote dataset mean and standard deviation, while Eq. (1) uses M and V for label component mean and variance; this notation conflict makes the preprocessing and decomposition steps harder to follow.","section":"Eq. (12) and Eq. (1)"},{"comment":"The caption says 'SIX BENCHMARK DATASETS' but the table reports results on three wireless traffic datasets; the caption should be corrected or clarified.","section":"Table I caption"},{"comment":"The text states that batch size is set to 1 for large-scale graph WT datasets, but the interaction between RSS subgraph sampling and batching is not explained; it is unclear whether each batch is one subgraph or one time window across all subgraphs.","section":"Section IV.B"},{"comment":"The MSE column is reported without specifying the dataset and prediction setting; it appears to correspond to Milano input-36-predict-24, but this should be stated explicitly in the table or caption.","section":"Table IV"},{"comment":"The caption mentions 'Flops (Bubble Size)' but gives no legend or numerical scale for bubble sizes, making the efficiency comparison difficult to interpret.","section":"Fig. 1 caption"},{"comment":"WTFlib is described as an open-source library, but no formal citation, version number, or documentation reference is provided; the anonymous link contains a visible username, so the submission is not anonymized.","section":"Abstract and Section I"}],"recommendation":"major_revision","confidential_remarks":"The load-bearing empirical claim in the abstract is not supported by the reported table, and the missing Appendix II is a serious omission. I recommend major revision: the authors should recompute the average improvements against the strongest baseline per dataset, add error bars or multiple-seed results, provide the proof of Theorem 1, and either implement a genuinely progressive training schedule or rename the method. If the 2% C2TM figure cannot be substantiated, the paper should be rejected rather than accepted with minor changes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core idea—decompose the target into cross-sectional mean/variance or STL components, supervise separate branches on each, then recombine—is fresh, and the decomposition math is exact. The RSS subgraph sampling is a practical engineering answer to city-scale graphs on a single GPU, and WTFlib is a real contribution. But the paper's central empirical claims are overstated, and the evidence as reported is not trustworthy enough to take the 2/4/11% numbers at face value.\n\nWhat's genuinely new: most decomposition work in forecasting decomposes inputs; decomposing the labels and adding per-component losses is a reasonable, relatively unexplored trick. The MVD decomposition is a deterministic transform of the target, so no circularity worry there. The ablations show component losses help a bit, and the framework is model-agnostic, which is nice.\n\nThe soft spots are real, and the stress-test note is right. Averaging Table I, PSLD-MVD beats the best baseline on C2TM by about 0.06%, not 2%. The 11% on CBS only appears against PatchTST; against TimeMachine it's roughly 4.8%. No error bars or seeds are reported, and some numbers look unreviewed—FourierGNN's CBS MSE drops from 2.601 at horizon 24 to 1.193 at horizon 72, which is implausible. Theorem 1's proof is cited as Appendix II, which isn't in the text, and Algorithm 1 doesn't actually show subgraph partitioning. Also, 'progressive supervision' is a misnomer: the components are learned with static auxiliary losses, not a curriculum. That's a terminology issue, but it matters because the paper sells the progressiveness as the mechanism.\n\nThe math itself is fine. The recombination in Eq. 5 is approximate but reasonable; the identity-recombination case is exact. The RSS theorem may be true, but without a proof we can't evaluate it.\n\nWho this is for: someone working on multivariate wireless traffic forecasting or large-graph training will get ideas here, and WTFlib is useful. But the paper needs major revision: fix the aggregates, add error bars, supply the missing proof, and either rename or reframe the 'progressive' claim.\n\nRecommendation: send it to peer review. A serious referee can sort out the empirical issues. If I were the editor, I'd treat the current claims as provisional and ask for a revised version before acceptance.","headline":"Plausible label-decomposition trick with an overclaimed empirical story; worth refereeing, but the headline numbers don't survive contact with Table I.","tokens_in":19401,"tokens_out":2188,"would_cite":false,"duration_ms":20842,"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":"PSLD claims that decomposing the forecast target into mean, variance, and residual components and supervising each separately yields average error reductions of 2%, 4%, and 11% over thirteen baselines on three wireless traffic datasets.","keywords":["wireless traffic forecasting","long-term time series forecasting","large-scale graph sampling","label decomposition","random subgraph sampling","non-stationary time series","progressive supervision"],"falsifier":"Train PSLD with the component-loss weight $\\lambda$ set to zero but keep the same decomposition, branches, and combiner; if the accuracy gain over a plain MLP disappears, the improvement comes from the extra component supervision, and if it persists, the decomposition architecture alone is doing the work.","tokens_in":18303,"feed_emoji":"📡","tokens_out":8756,"duration_ms":80662,"temperature":0.7,"pith_summary":"Long-term forecasts of city-wide wireless traffic are hard because the series are non-stationary and the networks contain thousands of nodes. This paper argues that the right fix is to decompose the thing being predicted, not just the input: split each target series into a mean, a variance, and a normalized residual (or, in a second version, a trend, a seasonal part, and a residual), train a separate lightweight branch on each component, then recombine the branches with learned weights. To make training on very large graphs feasible, it samples random subgraphs at each step and claims these samples are unbiased stand-ins for the full graph. On three city-scale traffic datasets the method beats thirteen baselines with average error reductions of roughly 2%, 4%, and 11%, while running faster than most competitors.","feed_headline":"Forecast error drops 2-11% when the target is decomposed","feed_subtitle":"Splitting traffic series into mean, variance, and residual components beats 13 baselines on three city-scale datasets.","key_machinery":"The load-bearing object is the decomposer–learner–combiner chain. The Mean-Variance Decomposer computes $M = \\operatorname{Mean}(Y, -1)$, $V = \\operatorname{Var}(Y-M, -1)$, and a normalized residual $R = (Y-M)/V + \\epsilon$; the STL variant instead isolates trend $T$, seasonal $S$, and residual $R$. Each component is fed to its own learner and predictor, producing $\\hat M$, $\\hat V$, $\\hat R$, and the combinator reconstructs $\\hat Y = P_\\xi(L_\\xi(\\hat V \\hat R) + \\hat M)$ (or $\\hat Y = P_\\xi(L_\\xi(S + R) + T)$ for STL). The loss adds per-component prediction errors to the final prediction error, so gradients flow to the shallow layers from each component separately. Random Subgraph Sampling (RSS) is the supporting mechanism that makes this trainable at city scale: at each iteration it draws a random subset of nodes and the induced edges, and Theorem 1 asserts that across iterations the sampled subgraphs are unbiased estimators of the full graph's aggregate features.","core_discovery":"PSLD's central discovery is that treating decomposition as a source of supervision signals—rather than as input preprocessing or an attention module—lets a simple MLP beat much larger transformer and graph models on long-term wireless traffic forecasting. The target matrix is decomposed into components that are each statistically simpler, e.g. $M=\\operatorname{Mean}(Y)$, $V=\\operatorname{Var}(Y-M)$, $R=(Y-M)/V+\\epsilon$, and separate predictors are trained to match each component before a combiner reconstructs the final output. Theorem 1 states that subgraphs sampled by the Random Subgraph Sampling algorithm are an unbiased estimator of the true aggregate features of the whole graph, which is the paper's justification for training on subgraphs. The reported result is best accuracy among the compared models across all tested prediction lengths, with average MSE improvements of about 2% on C2TM, 4% on Milano, and 11% on CBS.","pith_inferences":["The label 'progressive' implies a curriculum, but the loss function sums all component losses from the start; whether ordering or gating component supervision would change results is an untested extension, and the framing predicts it should matter.","Because the variance branch is trained to match the true spread of the series, it could be calibrated into a prediction interval; the paper notes this as future work rather than demonstrating it.","The proof of Theorem 1 is referenced as Appendix II, but that appendix does not appear in the posted version; the unbiasedness claim should be read as asserted until the proof is available.","RSS's unbiasedness is stated for aggregate features, not for edge-level topology; plugging graph-aware backbones into PSLD would reveal whether random node subgraphs preserve enough structure for spatial learning."],"forward_implications":["City-scale long-term forecasting becomes feasible on a single GPU: random subgraph sampling keeps each batch small while iterating over the whole graph across training steps.","Forecasters become more interpretable: the mean, variance, and residual heads can be inspected separately, and the variance head offers a natural way to see where the model expects more spread.","The mechanism is not tied to the MLP backbone; adding PSLD-style supervision to PatchTST, STID, and GWNet improved their accuracy in the paper's experiments.","The benefit generalizes beyond wireless traffic: PSLD also reports lower MSE than strong baselines on ETT, ILI, Weather, Electricity, and Traffic benchmarks."],"supporting_citations":[{"why":"Supplies the STL seasonal-trend decomposition used as the second decomposer and the decomposition vocabulary of trend/seasonal/residual.","marker":"[35]"},{"why":"Input-decomposition Transformer baseline whose decomposition-in-module approach PSLD contrasts with and outperforms.","marker":"[31]"},{"why":"Frequency-enhanced decomposed Transformer baseline used in the comparisons.","marker":"[32]"},{"why":"PatchTST is one of the strongest baselines; PSLD reports relative MSE reductions against it.","marker":"[45]"},{"why":"DLinear provides the fast linear baseline and the comparison point for runtime-versus-accuracy trade-offs.","marker":"[48]"},{"why":"Informer supplies the long-sequence transformer baseline and the ETT benchmark protocol used in generalization tests.","marker":"[41]"},{"why":"Provides the Milano dataset, one of the three large-scale wireless traffic benchmarks.","marker":"[39]"},{"why":"Provides the C2TM dataset, the city-scale Chinese wireless traffic benchmark.","marker":"[38]"}],"fun_headline_variants":["Decompose the target to beat transformers on traffic forecasting","Progressive label decomposition lifts wireless traffic forecast accuracy","Subgraph sampling plus decomposed targets improves traffic forecasts","Splitting target into components boosts long-term traffic forecast","Why decompose? It turns simpler predictors into SOTA forecasters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's load-bearing bet is that splitting each output series into mean, variance, and normalized residual (or trend, seasonal, and residual) yields components that are each easier to predict, and that supervising them separately before recombining improves the final forecast enough to justify the extra machinery.","fun_headline_variants_meta":{"raw":{"variants":["Decompose the target to beat transformers on traffic forecasting","Progressive label decomposition lifts wireless traffic forecast accuracy","Subgraph sampling plus decomposed targets improves traffic forecasts","Splitting target into components boosts long-term traffic forecast","Why decompose? It turns simpler predictors into SOTA forecasters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1336,"prompt_tokens":991,"completion_tokens":345,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":281}},"tokens_in":607,"tokens_out":345,"duration_ms":4158,"temperature":1.0,"reasoning_tokens":281,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:21:18.082023+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train PSLD with the component-loss weight $\\lambda$ set to zero but keep the same decomposition, branches, and combiner; if the accuracy gain over a plain MLP disappears, the improvement comes from the extra component supervision, and if it persists, the decomposition architecture alone is doing the work.","supporting_citations":[{"cited_title":"Stl: A seasonal-trend decomposition,","cited_arxiv_id":null,"evidence_quote":"Supplies the STL seasonal-trend decomposition used as the second decomposer and the decomposition vocabulary of trend/seasonal/residual."},{"cited_title":"Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting,","cited_arxiv_id":null,"evidence_quote":"Input-decomposition Transformer baseline whose decomposition-in-module approach PSLD contrasts with and outperforms."},{"cited_title":"FED- former: Frequency enhanced decomposed transformer for long-term series forecasting,","cited_arxiv_id":null,"evidence_quote":"Frequency-enhanced decomposed Transformer baseline used in the comparisons."},{"cited_title":"A time series is worth 64 words: Long-term forecasting with transformers,","cited_arxiv_id":null,"evidence_quote":"PatchTST is one of the strongest baselines; PSLD reports relative MSE reductions against it."},{"cited_title":"Informer: Beyond efficient transformer for long sequence time-series forecasting,","cited_arxiv_id":null,"evidence_quote":"Informer supplies the long-sequence transformer baseline and the ETT benchmark protocol used in generalization tests."},{"cited_title":"A multi-source dataset of urban life in the city of milan and the province of trentino,","cited_arxiv_id":null,"evidence_quote":"Provides the Milano dataset, one of the three large-scale wireless traffic benchmarks."},{"cited_title":"Analyzing and modeling spatio-temporal dependence of cellular traffic at city scale,","cited_arxiv_id":null,"evidence_quote":"Provides the C2TM dataset, the city-scale Chinese wireless traffic benchmark."}],"review_version":1}