{"id":"9ed4da78-2f40-46b5-ae09-f1144ce01940","arxiv_id":"2501.13794","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"NPDiff blends a data-dynamics-derived noise prior with the learned denoiser in diffusion models, reporting large accuracy gains on mobile traffic forecasting tasks.","lead":"This paper adds a simple noise prior to diffusion models for mobile traffic prediction, injecting seasonal and short-term patterns directly into the denoising step. The authors report accuracy gains of over 30% compared with a diffusion baseline on four city-scale cellular traffic datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be inflated by selecting λ and NK on the test set; validation-based hyperparameter selection could shrink or erase the claimed 30% improvement.","rationale":"The reader's weakest assumption was the theoretical correctness of the modified reverse sampling in Eq. (14). That is a real gap: the paper provides no proof that the weighted noise estimate bϵ = λeϵ + (1−λ)ϵθ samples the correct predictive distribution, and the training objective in Eq. (13) is heuristic. However, the central claim is empirical: NPDiff 'delivers predictions that are effective, efficient, and robust' and 'achieves superior performance with an improvement over 30%.' Even if the sampling theory were fully settled, the reported gains would still be suspect if the hyperparameters were tuned on the test set. Appendix A.2 is ambiguous: it says the best performance is reported from all combinations of NK and λ without specifying the split used for selection, and no error bars are given anywhere in Tables 2, 3, or 5-9. This is a falsifiable methodological flaw that can directly invalidate the headline improvement. The weak CSDI baseline further suggests the 'over 30%' figure is against a poorly tuned diffusion model, not against the state of the art. A re-run with validation-only selection and proper error bars is a concrete, decisive check. Since the paper could be repaired by such a re-run, the conditional-accept verdict stands, but the reason for conditionality should focus on the evaluation protocol rather than only on the missing sampling proof.","tokens_in":19855,"tokens_out":5528,"duration_ms":50261,"concrete_test":"Re-run the full comparison with a strict validation protocol: select NK and λ on the 20% validation split using the paper's early-stopping rule (3 sampled runs), freeze the selected hyperparameters, and report test MAE/RMSE over 5 seeds as mean ± std. Compare CSDI+Prior against CSDI and against the strongest non-diffusion baselines (PromptST, STNorm, TAU). If the validation-selected configuration does not preserve a >30% MAE reduction over CSDI, or if the gain over the best non-diffusion baseline becomes non-significant, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is the 'improvement over 30%' in the Abstract, supported by Tables 2, 3, and the appendix. Appendix A.2 describes the model hyperparameters: NK is chosen from {2, 3, 5, 7, N_mean} and λ from {0.3, ..., 0.7}, and then states 'we report the best performance selected from all combinations of these two parameter settings.' The paper does not explicitly say that this selection is performed on the validation split. If the best combination is chosen on the test split, the reported errors are optimistically biased, and the comparison against CSDI and all baselines is not a fair estimate of generalization. This concern is load-bearing because the main evidence for NPDiff is the empirical tables; without a clean validation protocol, the headline 30-40% improvement cannot be distinguished from overfitting to the test split. The concern is amplified by the anomalously weak CSDI baseline (e.g., MobileBJ CSDI MAE 0.596 vs HA 0.232), so a large fraction of the gain may reflect repairing a misconfigured baseline rather than the noise prior itself. A re-evaluation with validation-only hyperparameter selection and error bars would settle whether the claimed superiority is real.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NPDiff, a plug-in framework for diffusion-based mobile traffic forecasting. The central idea is to decompose the diffusion noise ϵ into a data-dynamics-derived \"noise prior\" eϵ and a residual Δϵ, where eϵ is computed from periodic dynamics (FFT-based, Eq. 4-5) for multi-step prediction and from local dynamics (lag-1 values, Eq. 6) for one-step prediction. The final noise estimate is a weighted blend bϵ = λeϵ + (1−λ)ϵθ of the prior and the denoising network output (Eq. 12). The training objective is modified to predict ϵ against bϵ (Eq. 13), and the reverse update replaces ϵθ with bϵ (Eq. 14). Experiments cover four mobile traffic datasets, three diffusion backbones (CSDI, ConvLSTM, STID), multi-step and one-step tasks, and report an average MAE improvement over the unmodified CSDI baseline of about 41.6% in the 12-12 task, supporting the Abstract's claim of \"improvement over 30%.\" The paper also reports faster early convergence and reduced prediction uncertainty with the noise prior.","tokens_in":20092,"tokens_out":3328,"duration_ms":33262,"significance":"If the reported gains are real, the paper makes a useful conceptual contribution: it identifies that the noise being estimated in a diffusion forecaster can be regularized with data-driven structure, and it demonstrates gains across several denoising architectures. The framework is simple, the derivation in Section 5.3 is algebraically coherent, code and data are promised, and the ablation on λ and NK provides some insight into sensitivity. However, the paper's central quantitative claim rests on three issues that need to be addressed before the result can be accepted at face value: the hyperparameters λ and NK appear to be selected on the test split rather than on a clean validation protocol; the empirical tables have no error bars or significance tests; and the CSDI baseline is anomalously weak on some datasets, so a large fraction of the reported improvement may be repairing a misconfigured baseline. The theoretical guarantee for the modified sampling procedure is also not established, and the uncertainty reduction in Figure 8 is plausibly a shrinkage artifact rather than evidence of better probabilistic calibration.","major_comments":[{"comment":"Appendix A.2 states that for the model's hyperparameters, NK is chosen from {2,3,5,7,N_mean} and λ from {0.3,...,0.7}, and \"we report the best performance selected from all combinations of these two parameter settings.\" The paper never states that this selection is performed on the validation split. If the selection is made on the test set, all numbers in Tables 2, 3, and 5-9 are optimistically biased, and the headline \"improvement over 30%\" cannot be distinguished from test-set overfitting. The authors must specify explicitly that selection was on validation data, and ideally report the selected λ and NK per dataset per task. If the selection was on test data, the experiments must be redone with validation-only selection.","section":"Appendix A.2"},{"comment":"The evaluation procedure samples 50 times at test time and reports the average, but no standard deviations, confidence intervals, or significance tests are reported anywhere in the tables. Because diffusion sampling is stochastic and the reported differences are often small (e.g., STID+Prior vs STID on MobileSH14 in Table 5, MAE 0.037 vs 0.046; CSDI+Prior vs CSDI on MobileSH14 in Table 2), the reader cannot determine whether the differences are statistically reliable. The paper should report error bars over the 50 samples or over multiple training seeds, and where differences are small, run paired significance tests. The 90% confidence intervals shown qualitatively in Figure 8 are not sufficient and are not reported numerically for the headline tables.","section":"Tables 2, 3, 5-9"},{"comment":"The CSDI baseline is anomalously weak in several settings: on MobileBJ, CSDI achieves MAE 0.596 and RMSE 3.178, while the simple History Average baseline achieves MAE 0.232 and the STID backbone achieves MAE 0.116; even after adding the prior, CSDI+Prior drops to 0.094. Such a large gap indicates that the CSDI implementation is likely undertuned or misconfigured for this task, and the paper's headline improvement of 41.6% in MAE is computed against this weak baseline. The authors should either tune CSDI with the same care as the baselines, include a stronger diffusion baseline (e.g., DiffSTG or a recent time-series diffusion model), or at minimum report the configuration used for CSDI and show that its poor MobileBJ performance is not due to a fixable implementation issue.","section":"Table 2"},{"comment":"The noise prior eϵ in Eq. (10) is obtained by substituting the extracted dynamics D for the target x_0^ta in the rearranged forward noising equation (Eqs. 7-9). This means the reverse update in Eq. (14) is, by construction, biased toward D when λ > 0. The paper does not provide any analysis of what distribution the modified reverse chain samples, nor does it show that the resulting predictive distribution is calibrated. The reduced uncertainty in Figure 8 may therefore simply reflect shrinkage toward a deterministic seasonal/lag-1 prior rather than an accurate probabilistic model. As a concrete test, the authors should compare NPDiff against a simpler post-hoc procedure that blends the CSDI predictive mean with D (or uses D as an additional conditioning feature) using the same λ. If the simple procedure matches NPDiff's gains, the claim that the noise prior specifically, rather than target shrinkage, is responsible for the improvement would not be supported.","section":"Section 5.3, Eqs. (7)-(14)"}],"minor_comments":[{"comment":"The noise schedule is written as {β_n}_{N}^{t=1}; the index should be n=1,...,N. This is likely a typo but should be corrected.","section":"Eq. (1)"},{"comment":"There are LaTeX artifacts in the text: \"decomposes noise intoprior\" (missing space) and \"prior}\" with a stray brace in the Abstract. These should be cleaned before publication.","section":"Abstract and Section 1"},{"comment":"The axis labels in Figures 4 and 5 render as corrupted Unicode strings such as \"/uni00000013/uni00000011...\" instead of readable text. The figures need to be regenerated.","section":"Figures 4 and 5"},{"comment":"The Time-LLM reference [Jin et al.] has no year or venue; please complete the citation.","section":"References"},{"comment":"The notation for the number of FFT components is inconsistent: the paper uses NK in Section 5.2 and Appendix A.2, but writes \"Nk\" in the opening of Section 6.3. Please unify.","section":"Section 6.3"}],"recommendation":"major_revision","confidential_remarks":"The most important issue to verify editorially is the hyperparameter-selection protocol in Appendix A.2. If λ and NK were selected on the test set, the headline 30% improvement is likely optimistic and the paper should be asked to rerun with validation-only selection. The anomalously weak CSDI baseline on MobileBJ also suggests that the comparison may be unfair; the authors should provide the exact CSDI configuration and ideally add a stronger diffusion baseline. The theoretical concern about the modified sampling procedure in Eq. (14) is a correctness-risk issue, but it can be addressed experimentally with a simple post-hoc blending baseline, so I do not view it as grounds for rejection if the empirical claims are re-evaluated properly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: NPDiff is a simple and plausible idea—decompose the diffusion noise into a prior derived from data dynamics (FFT-based seasonal pattern for multi-step, lag-1 value for one-step) and a residual learned by the denoising network, then blend with λ. The algebra in Section 5.3 is correct; the novelty is the application of the noise-prior trick from video diffusion to mobile traffic, with dynamics chosen to match that data. The paper is honest about the derivation being a rearrangement.\n\nWhat the paper does well: It runs a broad evaluation—four real cellular traffic datasets, three diffusion backbones (CSDI, ConvLSTM, STID), and multiple horizons. The improvements over the vanilla diffusion versions are large and consistent, and NPDiff+CSDI also beats strong non-diffusion baselines like iTransformer, Time-LLM, and PromptST on most datasets. They show faster convergence and less uncertainty. Code and data are promised. That is real work, not a toy.\n\nWhere I'd push back, in proportion:\n\nFirst, the stress-test note is right. Appendix A.2 says the reported results are 'the best performance selected from all combinations' of NK and λ, but it never says that selection was on the validation split. If it was test-set selection, the headline 30% is optimistically biased. This is the single most important thing to fix. The authors need to state the split explicitly and, ideally, re-run with validation-only selection.\n\nSecond, the CSDI baseline is surprisingly weak—MobileBJ CSDI MAE 0.596 vs HA 0.232 suggests the vanilla diffusion model is undertuned. That inflates the relative gain. I'd ask for a properly tuned CSDI or at least an explanation.\n\nThird, no error bars or significance tests. Since they already sample 50 times at test, reporting standard deviations would be cheap and would make the comparisons much more convincing.\n\nThe absence of a formal guarantee for the modified reverse process (Eq. 14) is a real limitation, but not fatal. The method is heuristic; the paper doesn't overclaim theory. If D is wrong, the prior biases predictions, and the paper's cosine-similarity analysis (Table 1) is the only guard.\n\nWho it's for: people working on diffusion for spatio-temporal forecasting, especially mobile/crowd traffic. It's a practical plug-in with likely genuine value, and it's a useful case study in how evaluation choices can drive reported gains.\n\nRecommendation: send to peer review, but as a conditional accept. Ask for (1) clarification and validation of hyperparameter selection, (2) error bars, and (3) a retuned CSDI baseline. The paper deserves a serious referee.","headline":"Simple noise-prior plug-in that likely helps, but the headline 30% needs validation-only hyperparameters and error bars before I'd trust it.","tokens_in":20630,"tokens_out":3268,"would_cite":true,"duration_ms":29341,"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":"Decomposing diffusion noise into a data-dynamics prior and a residual improves mobile traffic prediction by over 30%, while speeding training and tightening uncertainty.","keywords":["mobile traffic prediction","diffusion models","noise prior","periodic dynamics","local dynamics","spatio-temporal forecasting","denoising network","cellular network traffic"],"falsifier":"Deliberately break the training periodicity in the test set with a public holiday, network outage, or shifted weekday schedule, and compare NPDiff at a high $\\lambda$ against the same model at $\\lambda = 0$; if the prior-anchored model degrades below the no-prior baseline in such periods, the assumption that $D$ tracks $x^{ta}_0$ has failed.","tokens_in":19630,"feed_emoji":"📡","tokens_out":6001,"duration_ms":50408,"temperature":0.7,"pith_summary":"Mobile traffic from base stations is strongly periodic and locally autocorrelated, and this paper claims that the noise a diffusion model learns to remove carries those same patterns. The proposed NPDiff framework splits the target noise into a prior, computed directly from extracted periodic and lag-1 dynamics, and a residual that the denoising network still predicts. A weighted blend of the two replaces the pure network output during training and sampling. On four real cellular-traffic datasets, across three different denoising backbones and several forecast horizons, this raises accuracy by more than 30% on average, speeds up early training, shrinks prediction uncertainty, and makes forecasts more robust to input noise. If correct, it means diffusion forecasting for traffic-like time series can be improved simply by engineering the noise, without redesigning the denoising network.","feed_headline":"Noise priors from traffic rhythms cut diffusion forecast error 30%","feed_subtitle":"Baking weekly periodicity and lag-1 dynamics into the denoising noise boosts accuracy, speed, and robustness.","key_machinery":"The load-bearing object is the noise prior $\\tilde{\\epsilon}$: a closed-form expression that produces a reference noise at every diffusion step from the data dynamics $D$, so that the network only has to learn the leftover. $D$ is chosen from two cheap statistics—a small set of dominant FFT components averaged over the weekly period for multi-step tasks, and the lag-1 previous value for one-step tasks. Fusing via $\\hat{\\epsilon} = \\lambda\\tilde{\\epsilon} + (1-\\lambda)\\epsilon_\\theta$ with $\\lambda \\approx 0.5$, the prior anchors the denoising trajectory while the residual keeps it responsive to irregular changes. This is what lets the method be bolted onto CSDI, ConvLSTM, and STID without changing their architectures.","core_discovery":"The central claim is that the Gaussian noise estimated during diffusion denoising is not pure randomness but a carrier of the data's own dynamics, and that exploiting this structure improves prediction. Concretely, with $D$ the extracted dynamics (FFT-based weekly periodicity for multi-step forecasting, lag-1 values for one-step forecasting), the target is written $x^{ta}_0 = D + \\Delta x$ with small $\\Delta x$. Substituting into the diffusion forward equation gives the noise as a computable prior $\\tilde{\\epsilon} = (x^{ta}_n - \\sqrt{\\bar\\alpha_n} D)/\\sqrt{1-\\bar\\alpha_n}$ plus a residual $\\Delta\\epsilon$. NPDiff trains the network against the blended estimate $\\hat{\\epsilon} = \\lambda\\tilde{\\epsilon} + (1-\\lambda)\\epsilon_\\theta$ and uses the same blend in the reverse update. The paper claims this yields the reported gains while slotting into existing denoising architectures unchanged.","pith_inferences":["The paper does not test this, but the same decomposition should transfer to other strongly seasonal and autocorrelated flow data such as energy load, water demand, or web traffic.","Because the optimal mixing weight lands near 0.5, the learned residual and the data-dynamics prior carry comparable weight; a step-dependent or dataset-adaptive $\\lambda$ may extract further gains, which the paper leaves unexplored.","A useful check the paper leaves open is to compare against a deterministic model that uses the same periodic and lag-1 dynamics directly, to see how much of the gain comes from the diffusion sampling loop rather than from the dynamics themselves.","Under distribution shift—holidays, lockdowns, or cell-site changes—the prior extracted from training data can go stale, so an online-updated $D$ or an uncertainty-gated $\\lambda$ would be a natural next step."],"forward_implications":["Any existing diffusion-based traffic predictor can adopt the noise prior as a plug-in, so the reported gains are not tied to a specific denoising architecture.","Multi-step forecasting benefits most from the periodic prior, while one-step forecasting benefits most from the lag-1 prior, meaning the choice of dynamics should match the task horizon.","Because part of the noise is computed rather than learned, the model converges faster early in training, which matters for real-time resource management.","Noise-prior predictions are more stable under injected Gaussian perturbation, indicating greater robustness in noisy operational data.","Uncertainty estimates tighten substantially, which improves the reliability of probabilistic forecasts for network operators."],"supporting_citations":[{"why":"Defines the forward/reverse diffusion process and the noise-prediction objective that NPDiff modifies.","marker":"[Ho et al., 2020]"},{"why":"Supplies CSDI, the transformer-based diffusion backbone that NPDiff wraps and compares against.","marker":"[Tashiro et al., 2021]"},{"why":"A representative conditional diffusion model for time series whose conditioning approach NPDiff extends to the noise component.","marker":"[Shen and Kwok, 2023]"},{"why":"Shows that a noise prior can preserve correlation in video diffusion, motivating the idea that noise carries structure.","marker":"[Ge et al., 2023]"},{"why":"Cited directly as the basis for the assumption $x_0 = D + \\Delta x$ and for using a static reference signal as a noise prior.","marker":"[Zhang et al., 2024]"},{"why":"A temporally correlated noise prior for video diffusion that this paper builds on conceptually.","marker":"[Chang et al., 2024]"},{"why":"Provides ConvLSTM, one of the three denoising backbones used to demonstrate the framework's generality.","marker":"[Shi et al., 2015]"},{"why":"Provides STID, one of the three denoising backbones used to demonstrate the framework's generality.","marker":"[Shao et al., 2022]"}],"fun_headline_variants":["Noise priors cut diffusion traffic error 30%","Diffusion noise itself encodes traffic patterns: 30% gain","Noise priors boost diffusion forecasts by 30%","NPDiff uses noise dynamics for 30% better traffic prediction","Traffic cycles hidden in noise yield 30% diffusion gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the periodic and one-step-lag patterns measured from the training period are a good stand-in for the specific future values being predicted, so that the leftover $\\Delta x$ is small; when that fails, the noise prior would inject a strong and possibly harmful bias.","fun_headline_variants_meta":{"raw":{"variants":["Noise priors cut diffusion traffic error 30%","Diffusion noise itself encodes traffic patterns: 30% gain","Noise priors boost diffusion forecasts by 30%","NPDiff uses noise dynamics for 30% better traffic prediction","Traffic cycles hidden in noise yield 30% diffusion gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2859,"prompt_tokens":980,"completion_tokens":1879,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":1794}},"tokens_in":596,"tokens_out":1879,"duration_ms":13457,"temperature":1.0,"reasoning_tokens":1794,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:35:24.179330+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deliberately break the training periodicity in the test set with a public holiday, network outage, or shifted weekday schedule, and compare NPDiff at a high $\\lambda$ against the same model at $\\lambda = 0$; if the prior-anchored model degrades below the no-prior baseline in such periods, the assumption that $D$ tracks $x^{ta}_0$ has failed.","supporting_citations":[{"cited_title":"Csdi: Conditional score- based diffusion models for probabilistic time series impu- tation","cited_arxiv_id":null,"evidence_quote":"Supplies CSDI, the transformer-based diffusion backbone that NPDiff wraps and compares against."},{"cited_title":"Non- autoregressive conditional diffusion models for time se- ries prediction","cited_arxiv_id":null,"evidence_quote":"A representative conditional diffusion model for time series whose conditioning approach NPDiff extends to the noise component."},{"cited_title":"Trip: Temporal residual learning with image noise prior for image-to-video diffusion models","cited_arxiv_id":null,"evidence_quote":"Cited directly as the basis for the assumption $x_0 = D + \\Delta x$ and for using a static reference signal as a noise prior."},{"cited_title":"How i warped your noise: a temporally-correlated noise prior for diffusion models","cited_arxiv_id":null,"evidence_quote":"A temporally correlated noise prior for video diffusion that this paper builds on conceptually."},{"cited_title":"Convolutional lstm network: A machine learning approach for precipitation nowcasting","cited_arxiv_id":null,"evidence_quote":"Provides ConvLSTM, one of the three denoising backbones used to demonstrate the framework's generality."}],"review_version":1}