{"id":"7ddd4296-a3a0-45e4-b024-dcd5f6bf1260","arxiv_id":"2505.17431","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"HyperIMTS, a hypergraph neural network with temporal and variable hyperedges, achieves the lowest forecast MSE among 27 baselines on five irregular multivariate time series datasets.","lead":"HyperIMTS is a new hypergraph network for irregular multivariate time series forecasting that encodes each observation as a node and uses temporal and variable hyperedges for message passing without padding. Across five standard IMTS benchmarks with 27 baselines, it reports the lowest MSE in all five tables, though the text claims only four of five.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The equations after Eq. (2) never use the incidence matrices H_T/H_U, so the claimed hypergraph topology is not enforced; the reported gains may come from a global set-attention model rather than from irregularity-aware hypergraph message passing.","rationale":"The paper's central claim is that HyperIMTS achieves state-of-the-art results by leveraging irregularity-aware message passing in a hypergraph where the incidence structure ties observations to their own time and variable hyperedges. For that claim to hold, the forward pass must actually use that topology. As written, the formulation does not show this: H_T and H_U appear only in their definition and never in the message-passing equations. Eq. (6) is an unmasked attention over all observation nodes for both temporal and variable hyperedge updates; Eq. (14) is dimensionally underspecified and could be read as concatenating full hyperedge matrices rather than per-node gathered vectors. This is an internal-consistency issue, not a matter of consensus or tuning. If the code also ignores H, the method is a set/token attention architecture and the paper's explanation of its results is not established. I am not accusing the authors of misrepresentation: the equations may be compressed notation for a gather operation that the released code implements, and the code link is provided. That is exactly why a code inspection is the single decisive check. I also note that the reader's summary misreads Table 1: HyperIMTS is not best on USHCN, where Warpformer reports 0.1565 versus HyperIMTS 0.1738; the paper's 'four of five' wording is correct. The reader's benchmark-fairness concern remains valid and worth addressing, but the incidence-matrix gap is more load-bearing because it questions whether the proposed architecture is actually a hypergraph network. The conditional verdict is appropriate pending the code check, so I recommend no change to the reader's verdict.","tokens_in":20368,"tokens_out":15773,"duration_ms":134879,"concrete_test":"Inspect the released PyOmniTS code and trace HyperIMTS.forward(). For Eq. (6), check whether the attention for temporal hyperedge t masks or selects only nodes with timestamp t, and whether the variable-hyperedge update selects only nodes with variable u. For Eq. (14), check whether node j receives only e'_time(t_j) and e'_var(u_j) via incidence-based scatter/gather rather than the full E'_time and E'_var matrices. If neither selection occurs, the hypergraph incidence structure is unused and the paper's mechanism claim fails; if both occur, this concern is resolved and the benchmark-fairness question becomes the main remaining risk.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper defines incidence matrices H_T and H_U in Eq. (2) and states that observation nodes are connected only to their own temporal and variable hyperedges, but no subsequent equation uses these matrices. Eq. (6) updates each temporal hyperedge by softmax attention over all M observation nodes (keys V||E_var), with no mask selecting only nodes at the same timestamp; the analogous variable-hyperedge update is unmasked over all nodes as well. Eq. (12) attends over all variable hyperedges without using H_U, and Eq. (14) writes V'' = ReLU(V + FFnode(V'||E'_time||E'_var)) without specifying a per-node gather of only the connected temporal and variable hyperedge embeddings. If the implementation follows these equations literally, the hypergraph incidence structure is never enforced, and HyperIMTS reduces to a global set-attention network with timestamp and variable tokens. In that case the central claim that irregularity-aware message passing in a hypergraph drives the empirical gains is not supported by the architecture as written. The benchmark-fairness issue identified by the reader is real and could shrink the margin over GraFITi, but it presumes the method is what the paper describes; the incidence-matrix gap is more fundamental and internal to the formulation.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes HyperIMTS, a hypergraph neural network for forecasting irregular multivariate time series without padding. Observations are represented as nodes, and temporal and variable hyperedges connect them; the model alternates node-to-hyperedge attention, irregularity-aware hyperedge-to-hyperedge attention among variables, and hyperedge-to-node updates, followed by a linear decoder. The method is evaluated on five IMTS datasets in a unified benchmark against 27 baselines, reporting the lowest MSE on four datasets and up to 11.4% improvement over GraFITi, with additional ablations and efficiency analyses.","tokens_in":20611,"tokens_out":8418,"duration_ms":89489,"significance":"The paper's strengths are the breadth of the benchmark (27 baselines, five datasets, five seeds), the public code release, and the careful ablation isolating temporal/variable hyperedges and irregularity-aware dependencies. If the architecture is corrected to enforce the claimed hypergraph incidence structure, the hypergraph formulation is an appealing way to avoid padding and to model cross-variable dependencies without shared timestamps. The significance is conditional on the structural issue discussed below, since the equations as written do not enforce the claimed hypergraph topology.","major_comments":[{"comment":"The incidence matrices H_T and H_U defined in Eq. (2) are never used after their definition. Eq. (6) computes the updated temporal hyperedge with unmasked softmax attention over all M observation nodes, so a temporal hyperedge at time t receives messages from observations at every other timestamp; the analogous variable-hyperedge update is also unmasked over all nodes. Eq. (12) builds a complete attention graph among all U variable hyperedges, and Eqs. (14)–(15) concatenate the full hyperedge embeddings to every node rather than gathering only the hyperedges incident to that node. As written, the model is therefore a global set-attention network with timestamp and variable tokens, and the central claim that irregularity-aware hypergraph topology drives the empirical gains is not supported. Please either reintroduce masks/gathers based on H_T and H_U in Eqs. (6), (12), (14), and (15), or explicitly present and defend the global-attention interpretation.","section":"§4.1–§4.2, Eqs. (2), (6), (12), (14), (15)"},{"comment":"The statement 'For all classification models, we replace the final softmax layer with a linear layer to enable forecasting' is a substantial architectural modification for SeFT, mTAN, Raindrop, and Warpformer, whose original training objectives and hyperparameters are classification-oriented. No validation is provided that this linear-head adaptation is a strong or fair forecasting baseline, and the paper's conclusion of superiority over 27 state-of-the-art models depends on this protocol. Please justify the adaptation, report any tuning performed for the forecasting head, and consider comparing against published forecasting results or a tuned forecasting version of the strongest classification baselines.","section":"§A.4 (baseline details)"},{"comment":"The text states that α 'prioritize[s] Sobs over Svar if there are more aligned observations than unaligned ones,' but Eq. (11) sets α = Tshared/Ttotal for any Svar > δ and Sobs != 0, without requiring Tshared/Ttotal > 0.5. For a pair with Tshared/Ttotal = 0.2, α = 0.2 and Svar still dominates, contradicting the stated logic. Either add the condition Tshared/Ttotal > 0.5 to Eq. (11) or revise the explanation to reflect the actual weighting.","section":"§4.2.2, Eqs. (10)–(11)"}],"minor_comments":[{"comment":"The hypergraph definition 'E := ET ∩ EU' should presumably be a union, not an intersection, since temporal and variable hyperedges are disjoint sets; as written the notation is inconsistent with the subsequent text.","section":"§4.1"},{"comment":"'MIMIC-VI' is a typo for MIMIC-IV; the text also calls GraFITi the 'overall next best model,' which is true on average across the four datasets where HyperIMTS leads, but not on USHCN, where Warpformer and GRU-D have lower MSE than GraFITi.","section":"§5.2"},{"comment":"Appendix A.4.2 contains the typo 'numeber' for 'number'; Appendix A.3 contains the typos 'data volumn' for 'data volume' and 'grouth' for 'growth'.","section":"§A.4.2 and Appendix A.3"},{"comment":"The phrase 'making them learnable' is imprecise: Eq. (4) fixes sinusoidal encoding and applies a learnable linear map FF_time, so the temporal hyperedge embeddings are learnable only through that map.","section":"§4.2.1"}],"recommendation":"major_revision","confidential_remarks":"The structural gap in the equations should be checked against the released code before acceptance; if the code contains incidence-based masking that the equations omit, a revision can fix the paper quickly. The benchmarking fairness issue is also worth pressing, but it would not by itself justify rejection if the model's own contribution is accurately presented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: HyperIMTS gets strong numbers on five IMTS benchmarks, but the paper's central architectural claim—that message passing is restricted by hypergraph incidence—is not supported by the equations as written. The incidence matrices H_T and H_U are defined in Eq. (2) and then never used. Eq. (6) gives each temporal hyperedge attention over all M observation nodes, with no mask to nodes at that timestamp; the variable-hyperedge update similarly attends over all nodes. Eq. (14) feeds every node the full set of hyperedge embeddings. If the implementation matches the equations, this is a global set-attention model with temporal and variable tokens, not a hypergraph GNN. The ablations still show those tokens help, but the 'irregularity-aware hypergraph message passing' narrative is doing a lot of work.\n\nWhat's good: the non-padded representation is clean, the irregularity-aware similarity (Eqs. 8–11) with a learnable threshold is a sensible idea, and the benchmark is unusually broad—27 baselines across five datasets with five seeds, plus a released codebase. The efficiency comparisons are useful, and the paper is clearly written, which makes the incidence gap conspicuous.\n\nOn the reader's report: the 'four of five vs. Table 1' inconsistency is a misread. Warpformer gets 0.1565 on USHCN MSE, HyperIMTS 0.1738, so the text is accurate; HyperIMTS is second there. The baseline-fairness concern is real but secondary—using original-paper hyperparameters and adapting classification models by swapping the softmax head is a known limitation, and the reported margins over GraFITi could shrink with better tuning.\n\nBottom line: the empirical comparison is worth having, and the architecture idea is worth pursuing. But the paper as written does not demonstrate that the hypergraph topology is doing the work. A serious referee should ask for masks in the attention equations or an admitted reframing as set-attention, plus a sensitivity analysis on delta and baseline tuning. I'd send it to review: the benchmark alone is a service to the community, and the gap is fixable.\n\nFor a reading group, this is a good case study in how a well-specified model can be misdescribed.","headline":"A serious empirical benchmark with a clean non-padding model, but the paper's central claim that hypergraph incidence drives the results is not supported by the equations: the incidence matrices are defined and never used again.","tokens_in":21145,"tokens_out":4762,"would_cite":true,"duration_ms":34206,"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":"Irregular time series forecasting can skip padding by treating every observation as a hypergraph node and passing messages along time and variable hyperedges.","keywords":["irregular multivariate time series","hypergraph neural network","time series forecasting","message passing","irregularity-aware similarity","missing values","node prediction","efficient representation"],"falsifier":"Retune the strongest baselines per dataset, especially GraFITi and tPatchGNN, with a full hyperparameter search and end-to-end training of the classification-derived baselines on the same five benchmarks; if HyperIMTS no longer holds the lowest MSE on at least two datasets, the reported up-to-11.4% margin rests on the particular baseline configuration rather than the model.","tokens_in":20158,"feed_emoji":"📈","tokens_out":8449,"duration_ms":71049,"temperature":0.7,"pith_summary":"The paper sets out to show that irregular multivariate time series, where each variable is sampled at its own unaligned times, can be forecast accurately without padding the data into a regular grid. Its model, HyperIMTS, turns each observed value into a node in a hypergraph and links nodes with temporal hyperedges and variable hyperedges, so that messages flow among all observations through three stages of message passing. On five standard benchmarks it reports the lowest mean squared error, with up to 11.4% improvement over the best irregular-series competitor, while processing only the observed values rather than padded matrices. A reader should care because, if the claim holds, hypergraphs offer a single representation that captures both temporal and cross-variable dependencies directly from raw irregular samples.","feed_headline":"Hypergraph model cuts irregular-series forecast error by 11.4%","feed_subtitle":"With no padding, every observed value becomes a node passed along time and variable hyperedges.","key_machinery":"The load-bearing object is the irregularity-aware similarity blend that drives inter-variable message passing. For each pair of variables, a time-aware similarity $S_{obs}$ is computed from dot products of observation nodes that share timestamps, and an overall similarity $S_{var}$ from the variable hyperedge embeddings; the two are combined as $\\mathrm{S}_{IMTS} = \\alpha S_{obs} + (1-\\alpha)S_{var}$ with $\\alpha = T_{shared}/T_{total}$ when $S_{var} > \\delta$ and $S_{obs} \\neq 0$, and $\\alpha = 0$ otherwise, with $\\delta$ a learnable threshold initialized to 0.5. This lets the model use fine-grained aligned comparisons when variables are largely aligned and fall back to whole-series comparisons when they are not, a choice that is only possible because the hypergraph gives observation nodes a unified place where temporal and variable hyperedge messages can meet.","core_discovery":"HyperIMTS converts an irregular multivariate time series into a hypergraph where each observation is a node, temporal hyperedges connect observations sharing a timestamp, and variable hyperedges connect observations of the same variable. Forecasting becomes node prediction: the nodes to be forecast are zero-initialized and updated by three message-passing stages—node-to-hyperedge attention that refreshes temporal and variable hyperedge embeddings, hyperedge-to-hyperedge attention that passes messages between variables, and hyperedge-to-node updates that propagate both kinds of information back to observations. The distinctive mechanism is the irregularity-aware variable similarity, which blends a time-aware similarity computed only on time-aligned observation pairs with an overall series-level similarity between variable hyperedges, choosing the mix by the fraction of shared timestamps with a learnable threshold. The paper reports the lowest MSE on all five datasets in Table 1 (the text counts four, attributing the exception to USHCN's high variance) and up to 11.4% improvement over GraFITi.","pith_inferences":["The time-aware/overall similarity blend could be lifted out of the hypergraph and inserted into set-based or bipartite-graph forecasters, potentially giving those models a cheap way to handle partially aligned variables.","The node-prediction framing may transfer to imputation and classification tasks, since any masked node is already treated as a prediction target by zero-initialization.","The efficiency argument depends on observation count being much smaller than the padded grid; on nearly dense irregular series the attention costs would rise quadratically and the practical gap over padding methods would shrink.","A synthetic dataset with controlled alignment ratios (from fully shared to fully disjoint timestamps) could isolate exactly when the time-aware term pays off and when the fallback to overall similarity is what preserves performance."],"forward_implications":["Padding becomes unnecessary: HyperIMTS operates only on observed values, avoiding the data-volume growth that canonical and patch-aligned padding introduce.","Both temporal and variable dependencies are learned inside one hypergraph, so irregular forecasting reduces to a node prediction problem on that hypergraph.","The irregularity-aware similarity lets the model adaptively weight time-aligned versus whole-series comparisons, addressing partial alignment between variables.","On the five benchmarks the model reports the lowest MSE, up to 11.4% better than the next-best irregular-series model GraFITi, with lower computational cost than padding-based alternatives.","Because non-padding models keep efficiency roughly constant as lookback length grows, the approach scales better to long irregular windows than padding-based models."],"supporting_citations":[{"why":"GraFITi, the bipartite-graph baseline that is the strongest irregular-series competitor; also supplies the preprocessing setup reused for four of the five datasets.","marker":"Yalavarthi et al., 2024"},{"why":"tPatchGNN, a patch-aligned padding baseline whose patch settings are reused, and the source of the Human Activity preprocessing setup.","marker":"Zhang et al., 2024"},{"why":"Warpformer, a canonical padding baseline representing multi-scale warping approaches to irregular clinical time series.","marker":"Zhang et al., 2023a"},{"why":"SeFT, the set-based representation baseline, illustrating the alternative non-padding view of observations as unordered tuples.","marker":"Horn et al., 2020"},{"why":"mTAN, a multi-time attention network baseline that encodes time information through reference points.","marker":"Shukla & Marlin, 2020"},{"why":"Latent ODEs, the canonical continuous-latent padding approach that HyperIMTS contrasts with for efficiency.","marker":"Rubanova et al., 2019"},{"why":"GRU-D, an RNN baseline for multivariate time series with missing values, included among the IMTS methods.","marker":"Che et al., 2018"},{"why":"PatchTST, a regular multivariate forecasting baseline whose strong results motivate the unified comparison of regular and irregular models.","marker":"Nie et al., 2022"}],"fun_headline_variants":["Hypergraph turns irregular time series into nodes for better forecasting","No padding: hypergraph neural net forecasts irregular series with 11.4% less error","Irregular time series? HyperIMTS links observations via temporal and variable hyperedges","HyperIMTS: hypergraph message passing cuts irregular-series forecast error by 11.4%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The superiority over the twenty-seven baselines assumes those baselines are run close to their best settings under the paper's policy of reusing original hyperparameters and adapting classification models by replacing the final softmax layer with a linear layer.","fun_headline_variants_meta":{"raw":{"variants":["Hypergraph turns irregular time series into nodes for better forecasting","No padding: hypergraph neural net forecasts irregular series with 11.4% less error","Irregular time series? HyperIMTS links observations via temporal and variable hyperedges","HyperIMTS: hypergraph message passing cuts irregular-series forecast error by 11.4%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000201,"raw_usage":{"total_tokens":1369,"prompt_tokens":923,"completion_tokens":446,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":359}},"tokens_in":539,"tokens_out":446,"duration_ms":3576,"temperature":1.0,"reasoning_tokens":359,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:47:05.860965+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retune the strongest baselines per dataset, especially GraFITi and tPatchGNN, with a full hyperparameter search and end-to-end training of the classification-derived baselines on the same five benchmarks; if HyperIMTS no longer holds the lowest MSE on at least two datasets, the reported up-to-11.4% margin rests on the particular baseline configuration rather than the model.","supporting_citations":[{"cited_title":"Irregular Multivariate Time Series Forecasting : A Transformable Patching Graph Neural Networks Approach","cited_arxiv_id":null,"evidence_quote":"tPatchGNN, a patch-aligned padding baseline whose patch settings are reused, and the source of the Human Activity preprocessing setup."},{"cited_title":"Set Functions for Time Series","cited_arxiv_id":null,"evidence_quote":"SeFT, the set-based representation baseline, illustrating the alternative non-padding view of observations as unordered tuples."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"mTAN, a multi-time attention network baseline that encodes time information through reference points."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Latent ODEs, the canonical continuous-latent padding approach that HyperIMTS contrasts with for efficiency."},{"cited_title":"H., Sinthong, P., and Kalagnanam, J","cited_arxiv_id":null,"evidence_quote":"PatchTST, a regular multivariate forecasting baseline whose strong results motivate the unified comparison of regular and irregular models."}],"review_version":1}