{"id":"51af23e1-0a44-4535-b774-47671fb51614","arxiv_id":"2501.13041","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TimeFilter improves multivariate time series forecasting by dynamically filtering a patch-level spatial-temporal graph with a Mixture-of-Experts router, achieving state-of-the-art MSE on 13 benchmarks.","lead":"TimeFilter is a graph-based forecasting model that builds a network of patch-level relationships across channels and time, then uses a learned router to keep only the most useful relationships for each patch. It reports lower forecast errors than prior models on 13 public datasets, including weather, traffic, and energy data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The k-NN graph construction in Eq. (3) is the bottleneck: the unreported density factor alpha determines which dependencies the patch-specific router can ever select, and no sensitivity analysis isolates filtration gains from this hard pruning.","rationale":"I read the paper in good faith. The architecture is coherent, code is promised, error bars are given for the main long-term comparison, and the headline reductions in Section 5.2 are internally consistent with the full tables. The central claim is that patch-specific dynamic filtration of a spatial-temporal graph yields state-of-the-art forecasting accuracy. The weakest load-bearing link is not the MoE router or GNN aggregation; it is the fixed k-NN graph that precedes all filtration. Since alpha is never reported and no sensitivity analysis is given, the reader cannot tell whether the empirical gains come from the filtration idea or from a chosen graph sparsity. This does not make the paper invalid; it makes the claim conditional on an unreported hyperparameter. The reader's CONDITIONAL verdict already captures this, so I do not propose a verdict change. If the alpha sweep proposed above shows stable performance, the concern is resolved and the central claim stands.","tokens_in":28020,"tokens_out":14210,"duration_ms":153669,"concrete_test":"Run the released code and extract the per-dataset alpha used to produce Tables 1-3 (the repository is linked in the abstract). Then re-run Weather, Electricity, Traffic, Solar-Energy, and PEMS08 with alpha in {0.01, 0.05, 0.1, 0.2, 0.5, 1.0}, keeping the filtration module and all other settings fixed, and report averaged MSE/MAE plus the TimeFilter-versus-w/o-Filter gap and the rank versus Leddam/DUET at each alpha. If the SOTA ranking and the filtration gain are stable for all alpha >= 0.05 and match the submitted numbers at the reported alpha, the concern is resolved; if error varies by more than roughly 5% or the ranking changes, the k-NN density, not filtration, is the load-bearing component. Also report the missing Climate configuration from Table 6.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 constructs the spatial-temporal graph before any filtration: in Eq. (3), M = k-NN(GeLU(Dist(X_h)), alpha), with k = floor(alpha * n) and alpha described only as a hand-tuned scaling factor. No value for alpha appears in the main text, in Table 6, or in Appendix A.3, and no sensitivity study over alpha is reported. This is load-bearing because the Patch-Specific Filtration Module of Section 4.2 can only retain edges that survive this k-NN cut: the router chooses among temporal, spatial, and spatial-temporal regions of the already-pruned ego graphs G_i. If alpha is small, any useful dependency outside the k nearest patch neighbors is permanently removed and the filtration cannot recover it; if alpha is large, the graph is nearly complete and the hard pruning plays little role. The ablation w/o Filter in Table 11 keeps the same k-NN graph, so it isolates the filtration module conditional on an arbitrary alpha, not the adequacy of the graph itself. Consequently the reported SOTA gains in Tables 1-3 are under-specified: they could be driven by a favorable hand-tuned sparsity level rather than by patch-specific dynamic filtration. Verification requires the actual alpha used per dataset and a sweep over alpha.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TimeFilter, a multivariate time series forecasting framework with three modules: (i) a Spatial-Temporal Construction module that patches each channel and builds a k-NN graph over the C×N patch tokens, (ii) a Patch-Specific Filtration module that uses a noisy-gating MoE router with dynamic Top-p allocation to retain only the temporal, spatial, or spatial-temporal edges deemed relevant for each patch, and (iii) an Adaptive Graph Learning module that aggregates the filtered neighborhoods and projects to the forecast. The authors evaluate on 9 long-term and 4 short-term benchmarks and report state-of-the-art results, claiming a 4.48% MSE / 2.23% MAE reduction over Leddam at fixed look-back 96 and a 5.34% / 1.40% reduction over DUET/CCM under searched look-back horizons. The paper includes a Wilcoxon test against Leddam, error bars for that comparison, a six-way ablation of the filtering method, and a released code repository.","tokens_in":1857,"tokens_out":2550,"duration_ms":88211,"significance":"If the reported results are reproducible, TimeFilter offers a meaningful step beyond coarse-grained channel clustering: the patch-specific, dynamically routed filtration of spatial-temporal edges is a novel design in this literature, and the evaluation spans 13 datasets with diverse channel counts and frequencies. The paper deserves credit for releasing code, reporting standard deviations and a Wilcoxon test for the main comparison, and including full per-horizon tables in the appendix. The main weaknesses are that the central mechanism is under-specified (the k-NN sparsity factor α is never reported, loss weights λ1 and λ2 are missing, Top-p choices are incomplete) and that the optimal-horizon protocol and the ablation study are not documented tightly enough to attribute the gains to the proposed filtration rather than to favorable hyperparameter choices or evaluation settings. These are fixable in a revision.","major_comments":[{"comment":"The k-NN graph is constructed with k = floor(alpha * n), where alpha is described only as a hand-tuned scaling factor, and no value of alpha appears in the main text, Table 6, or Appendix A.3. This is load-bearing because the Patch-Specific Filtration Module (Section 4.2) can only select edges among those that survive this pruning: a too-small alpha permanently removes useful dependencies, while a too-large alpha makes the hard pruning negligible. The w/o Filter ablation in Table 11 retains the same k-NN graph, so it isolates the filtration module only conditional on an arbitrary alpha, not the adequacy of the graph itself. Please report alpha for every dataset and provide a sensitivity analysis, such as a sweep over alpha with and without filtration, to show that the claimed gains are due to patch-specific filtration rather than a favorable sparsity choice.","section":"4.1, Eq. (3); Table 6; Appendix A.3"},{"comment":"The optimal-horizon protocol is ambiguous. The text says the look-back length is searched from {192, 336, 512, 720}, but it does not state whether each model is evaluated at its own optimal L or at a single L chosen per dataset, for example the L that is optimal for TimeFilter. Table 9 reports results at all four prediction horizons but does not list which input length was selected for each model and dataset. This matters because the headline improvement of 5.34% and 1.40% over DUET and CCM is obtained under this protocol. Please specify the selection criterion, report the chosen L per model and dataset, and, for robustness, also provide results at a common L.","section":"5.2, Table 2; Appendix B.2, Table 9"},{"comment":"The loss weights lambda1 and lambda2 are defined but never reported for any dataset, and Top-p is said to be selected from {0.0, 0.5} but the per-dataset value is not given. These hyperparameters directly control the routing behavior in Eqs. (9)-(11) and the balance between the prediction loss and the auxiliary losses in Eqs. (16)-(17). Without these values, the experiments are not reproducible from the manuscript alone. Please include a complete per-dataset hyperparameter table that reports alpha, lambda1, lambda2, and Top-p, together with the existing entries in Table 6.","section":"4.4, Eq. (18); 5.1"},{"comment":"The ablation study reports single-run differences without error bars, and several differences are small relative to the reported standard deviations. For example, on Weather the MSE difference between TimeFilter and w/o Filter is 0.239 versus 0.244, compared with a reported plus-or-minus 0.006 standard deviation for TimeFilter in Table 7; on Traffic the difference is 0.407 versus 0.414. It is therefore not established that the filtration module, the dynamic allocation, or the auxiliary losses provide a statistically significant benefit. Please report mean plus-or-minus standard deviation over at least three seeds for the key ablation rows and, ideally, include a paired significance test between TimeFilter and the main competitors in the ablation table.","section":"5.3, Table 4; Appendix B.4, Table 11"}],"minor_comments":[{"comment":"Dist(X_h) is defined as a per-head tensor in R^{H x n x n}, but the text does not explain how the H heads are combined into the single n by n adjacency matrix M in Eq. (3); please state whether the heads are averaged, concatenated, or otherwise aggregated.","section":"4.1, Eqs. (2)-(3)"},{"comment":"The notation E^S in Eq. (6) lacks the subscript i used in Eqs. (4)-(5); please clarify the indexing.","section":"4.1, Eq. (6)"},{"comment":"The header contains the typos Catagories and Weahter; these should be Categories and Weather.","section":"Table 4"},{"comment":"In the Electricity row for prediction horizon 336, the MSE value is rendered as 0,175 with a comma instead of a decimal point; please correct it.","section":"Table 8"},{"comment":"The phrase we obtain the p-value of 4.66e-10, indicating a significant improvement at the 99% confidence level is imprecise; a p-value measures evidence under a significance level, so it should be rephrased as significant at the 1% significance level or similar, with the test procedure stated.","section":"5.2"},{"comment":"The contribution statement We for the first time advocate for a fine-grained segmentation of dependencies is an absolute novelty claim; consider softening it unless a systematic prior-art discussion is provided.","section":"1. Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the empirical design is broad, but the missing hyperparameter values (alpha, lambda1, lambda2, Top-p) and the ambiguous optimal-horizon protocol are essential for reproducibility and for validating the central attribution claim. The concerns are fixable within the manuscript's scope: the authors should provide the full configuration table, sensitivity analyses over alpha, and error bars on the ablations. I recommend a major revision and suggest the editor ask for a reproducibility checklist, since the code is available but the manuscript itself is not currently self-contained."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real contribution to dependency modeling for multivariate forecasting, and it deserves a proper referee, but the authors need to release the missing hyperparameters and clarify the look-back search before the exact SOTA numbers should be trusted.\n\nWhat's new: TimeFilter replaces coarse channel-clustering with patch-specific routing over three dependency types (temporal, spatial, spatial-temporal). The MoE-style dynamic expert allocation with Top-p is a sensible way to let each patch choose its own dependency subgraph. I don't know of another method that does this at patch granularity, so the architectural novelty is genuine. The evaluation is broad: 13 datasets, code released, and the main tables include standard deviations and a Wilcoxon test against the second-best baseline. The ablation shows the filtration module consistently beats no-filter and static alternatives, even if the margins are small (1-4% MSE on most datasets).\n\nSoft spots, in order of importance. First, the stress-test note is right: Eq. (3) builds the candidate graph with k-NN at sparsity alpha, and alpha never appears in the paper. The router can only keep edges that survive that cut, so without alpha the method is underspecified, and the reported gains are conditional on an arbitrary sparsity level. The code may contain the values, but the paper should state them. Second, Table 2's \"searched look-back horizon\" protocol is ambiguous: it reads as if TimeFilter's L was optimized per dataset, but it is not clear whether each baseline also got its own optimal L. If only TimeFilter got the search, the comparison is unfair. Third, lambda1 and lambda2 for the loss are never reported, and Top-p is only given as a set of candidates, not per dataset. These are smaller issues, but they make reproduction harder than it should be.\n\nNone of this is a fatal flaw. The central claim, that patch-specific dynamic filtration improves forecasting, holds up in the reported experiments and in the ablation. The weaknesses are incomplete reporting and one protocol ambiguity, both fixable in revision.\n\nWho should read it: anyone working on channel dependency modeling, GNNs for forecasting, or adaptive graph construction. It deserves a serious referee; I would recommend accept with revisions conditional on releasing alpha and the loss weights, and clarifying the look-back search.","headline":"A genuinely new patch-specific MoE filtration mechanism for time-series graphs, with a solid but under-specified experimental write-up; fix the missing hyperparameters and the look-back-search protocol and it deserves acceptance.","tokens_in":28824,"tokens_out":3446,"would_cite":true,"duration_ms":33468,"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":"TimeFilter claims that per-patch filtration of a spatial-temporal graph—keeping only the dependencies each time segment needs—outperforms channel clustering and full dependency modeling across 13 benchmarks.","keywords":["time series forecasting","graph neural networks","spatial-temporal graphs","patch-specific filtration","mixture of experts","channel dependency modeling","dynamic routing","long-term forecasting"],"falsifier":"Run TimeFilter on a dataset where a single known pair of channels carries all predictive signal, with that pair deliberately placed outside the k nearest neighbors at construction time; if the model still forecasts well, the filtration is not doing the work attributed to it. Alternatively, sweep the pruning factor α over a plausible range (for example 0.05 to 0.5) on Electricity or Traffic and re-run the best baseline under the same sweep: if the reported 4.48% gap shrinks or reverses for any α in that range, the claim that patch-specific filtration drives the gain is falsified.","tokens_in":27821,"feed_emoji":"📈","tokens_out":7555,"duration_ms":64361,"temperature":0.7,"pith_summary":"TimeFilter is a graph-neural-network framework for multivariate time series forecasting that treats every short segment (patch) of every channel as a node and learns, for each patch, which dependencies to keep. The paper argues that the two dominant strategies—channel-independent models that ignore inter-channel information and channel-dependent models that fuse all of it—are both wrong for real data, and that coarse channel clustering is too blunt because correlations change over time. TimeFilter constructs a spatial-temporal graph, then uses a mixture-of-experts router with dynamic Top-p allocation to select, per patch, a subset of temporal, spatial, and spatial-temporal edges, filtering out the rest. On 13 datasets spanning weather, electricity, traffic, and solar energy, the paper reports consistent state-of-the-art results, including a 4.48% MSE reduction over the second-best long-term baseline and a 13.54% MSE reduction on the PEMS08 short-term benchmark. The central claim is that patch-specific, dynamically filtered dependencies, rather than any fixed dependency policy, are what improve forecasting accuracy.","feed_headline":"Per-patch dependency filtering cuts forecast error 4.5%","feed_subtitle":"A GNN router keeps only the correlations each time segment needs, beating channel-clustering baselines on 13 datasets.","key_machinery":"The central object is the patch-specific ego-graph filtration: each of the n = C×N spatial-temporal patches becomes the center of an ego-graph whose edges are partitioned into three regions—temporal (same channel, different patches), spatial (different channels, same time), and spatial-temporal (everything else). A sparsely-gated mixture-of-experts router with dynamic expert allocation (Top-p cumulative confidence) selects a subset of the three filters per patch, producing a sparse adjacency matrix that is then fed to a GNN for aggregation and prediction. The k-NN construction of the initial graph (M = k-NN(GeLU(Dist(X_h)), α), with k = ⌊αn⌋) fixes the candidate edge set before filtration, so all later decisions are choices among those candidates.","core_discovery":"In the paper's own terms, TimeFilter establishes that dependency modeling in multivariate forecasting should be done at the patch level with per-patch filtration, not at the channel level with clustering or with a single global graph. The discovery is that decomposing the initial graph into patch-specific ego-graphs, dividing each into temporal, spatial, and spatial-temporal edge regions, and then letting a noisy-gated router assign one or more filters to each patch—so that the number and type of kept edges varies by patch—yields lower MSE and MAE than any fixed strategy across domains. The paper supports this with ablations showing that heuristic filters (Top-K, random-K, region-wise top-K, channel-wise filtering, or no filtering) all underperform the learned dynamic selection, and with a Wilcoxon test against the second-best baseline at the 99% confidence level.","pith_inferences":["If patch-specific filtration is the right inductive bias, then the categorical expert split into temporal/spatial/spatial-temporal may be less important than the per-patch sparsity level itself; a learned continuous edge mask with the same dynamic budget could be compared to isolate what carries the gain.","The k-NN ceiling implies that combining TimeFilter with a learnable or denser graph constructor (rather than a fixed α) would likely raise accuracy further; this is testable but goes beyond the paper's experiments.","The same ego-graph filtration idea transfers to other multivariate sequence tasks such as anomaly detection, imputation, and change-point detection, where the question is also which cross-channel correlations to trust at each moment.","Because dynamic Top-p allocation lets each patch choose a different number of experts, the architecture effectively implements a data-dependent sparsity schedule; one could test whether an explicit per-patch sparsity penalty replicates the results without the MoE machinery."],"forward_implications":["Patch-level filtration should replace channel-level clustering as the standard refinement step in multivariate forecasting, since the paper shows it beats CCM and DUET under matched settings.","Longer look-back horizons become usable without the usual noise penalty, because filtration removes spurious correlations; the paper demonstrates state-of-the-art results under searched horizons up to 720.","The same architecture generalizes across domains with very different correlation structures, from 7-channel ETT to 1763-channel Climate, without changing the dependency policy.","Short-term traffic forecasting, where spatial-temporal edges carry the signal, benefits most (13.54% MSE reduction on PEMS08), suggesting the method is especially valuable for sensor-network data.","Filtration performed per ego-graph is parallelizable, so the accuracy gain does not require a full dense-graph computation; the paper reports the efficiency is competitive with other GNN-, CNN-, and Transformer-based models."],"supporting_citations":[{"why":"Channel clustering baseline (CCM); the paper argues its coarse grouping misses time-varying dependencies that patch-specific filtration captures.","marker":"Chen et al., 2024"},{"why":"DUET baseline that extends clustering to the temporal dimension; the comparison target for the 5.34%/1.40% long-term gains under searched horizons and for short-term PEMS results.","marker":"Qiu et al., 2024b"},{"why":"Leddam, the second-best long-term baseline (L=96) whose MSE/MAE TimeFilter reduces by 4.48%/2.23%, and the Wilcoxon test comparison.","marker":"Yu et al., 2024"},{"why":"Source of the MoE-Dynamic expert allocation mechanism that TimeFilter adapts for deciding how many filters each patch needs.","marker":"Huang et al., 2024"},{"why":"CrossGNN, the channel-dependent baseline and the work that motivates resisting graph noise via ego-graph decomposition.","marker":"Huang et al., 2023"},{"why":"Scaling law for time series forecasting that justifies searching look-back horizons for the optimal input length in the second long-term experiment.","marker":"Shi et al., 2024"},{"why":"Sparsely-gated mixture-of-experts routing, the classic mechanism underlying TimeFilter's noisy gating and expert importance loss.","marker":"Shazeer et al., 2017b"}],"fun_headline_variants":["Per-patch graph filters cut time series forecast error 4.5%","Adaptive per-patch dependency pruning boosts forecasting","TimeFilter: learn which correlations each patch needs","GNN-driven patch filtering outperforms channel clustering","Fine-grained dependency filtering wins on 13 datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned per-patch filters can only keep or discard edges that survive an initial k-nearest-neighbor pruning of the graph, and the pruning factor α is hand-tuned and never reported; if that pruning deletes a dependency the forecast needs, no later stage can recover it.","fun_headline_variants_meta":{"raw":{"variants":["Per-patch graph filters cut time series forecast error 4.5%","Adaptive per-patch dependency pruning boosts forecasting","TimeFilter: learn which correlations each patch needs","GNN-driven patch filtering outperforms channel clustering","Fine-grained dependency filtering wins on 13 datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000304,"raw_usage":{"total_tokens":1713,"prompt_tokens":878,"completion_tokens":835,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":758}},"tokens_in":494,"tokens_out":835,"duration_ms":7483,"temperature":1.0,"reasoning_tokens":758,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:29:01.770994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TimeFilter on a dataset where a single known pair of channels carries all predictive signal, with that pair deliberately placed outside the k nearest neighbors at construction time; if the model still forecasts well, the filtration is not doing the work attributed to it. Alternatively, sweep the pruning factor α over a plausible range (for example 0.05 to 0.5) on Electricity or Traffic and re-run the best baseline under the same sweep: if the reported 4.48% gap shrinks or reverses for any α in that range, the claim that patch-specific filtration drives the gain is falsified.","supporting_citations":[],"review_version":1}