{"id":"6d6cd22e-fe86-492d-ab9e-fdd9d6fbefdf","arxiv_id":"2501.10917","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DecomposeWHAR, a decomposition-and-fusion model using depth-wise convolutions, Mamba, and self-attention, achieves state-of-the-art results on three wearable human activity recognition datasets.","lead":"This paper presents DecomposeWHAR, a neural network that splits wearable sensor signals by sensor, variable, and channel, then fuses them to recognize human activities. It reports large accuracy gains over prior methods on three standard benchmark datasets, with moderate computational cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on quoted baselines under possibly mismatched Realdisp/Skoda protocols, and Table 1's improvement deltas do not reconcile with the printed means.","rationale":"The reader identified the most load-bearing concern: the SOTA claim depends on comparing against baselines quoted from Miao et al. (2022), without demonstrated equivalence of preprocessing, user subset, or evaluation protocol. I agree with that assessment. The paper's ablations support internal design choices, but they do not establish external SOTA status; that comes only from Table 1. Treating quoted baselines as comparable requires identical data handling, which is not documented. The additional internal inconsistency in the printed improvement row strengthens the need to rerun baselines rather than accept quoted numbers. This does not imply the model architecture is flawed; it means the central claim is conditional on baseline comparability. The reader's CONDITIONAL verdict is therefore appropriate, and my stress-test does not move it.","tokens_in":11290,"tokens_out":10486,"duration_ms":111973,"concrete_test":"Re-run all Table 1 baselines using the released or official code under DecomposeWHAR's exact protocol: the same 10-user Realdisp subset, same sensor placements, same 800 ms window with identical stride and overlap, and same user splits, then recompute the Table 1 deltas from per-fold results. If DynamicWHAR under the identical protocol approaches DecomposeWHAR's Realdisp 96.64/96.10 numbers, or if correcting the printed deltas reduces the claimed margins by more than roughly 1 point, the stated SOTA advantage is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DecomposeWHAR beats all prior state of the art on three benchmarks. That claim rests entirely on Table 1, but the paper states that 'The results of other models are quoted from Miao et al. (2022)' and only HARMamba and IF-ConvTransformer were run by the authors. The datasets section says Realdisp uses only 10 of 17 users 'due to incomplete data,' and the evaluation section describes protocols such as leave-one-user-out and an 80/10/10 Skoda holdout with an 800 ms window. Nowhere does the paper establish that the quoted baselines used the same 10-user subset, the same sensor placement, the same window size and overlap, or the same folds. If Miao et al.'s Realdisp numbers were produced on a different subset or with different preprocessing, the reported 3–5 point margins could be evaluation artifacts rather than model quality. This concern is reinforced by internal arithmetic in Table 1: the printed DecomposeWHAR deltas do not match any visible runner-up difference. For example, the +6.93 macro-F1 on Opportunity cannot be reproduced from the rounded values (the largest visible gap is +5.91 versus DynamicWHAR, or +4.99 versus HARMamba), and several other deltas are off by 0.1–0.2 points. The comparison table therefore needs independent recomputation before the SOTA claim is accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DecomposeWHAR, a multi-sensor wearable human activity recognition model that decomposes input signals at sensor, variable, channel, and temporal levels and then fuses them hierarchically. The decomposition phase uses a modality-specific embedding and depth-wise convolutions to preserve per-variable temporal features; the fusion phase combines channel- and variable-level features, applies a Mamba block for global temporal aggregation, and uses a self-attention mechanism for cross-sensor interaction. The authors evaluate on Opportunity, Realdisp, and Skoda, reporting accuracy and macro-F1 improvements over several baselines, and they include ablations, parameter analyses, and an efficiency comparison.","tokens_in":11614,"tokens_out":3440,"duration_ms":35381,"significance":"If the reported gains are reproducible under a matched evaluation protocol, the paper would represent a meaningful step in multi-sensor WHAR: the architecture is well-motivated, the ablations are informative, and the efficiency measurements on a wearable device are a valuable addition. The authors also release code and provide standard deviations, which are good scientific practices. However, the central SOTA claim currently rests on baseline numbers quoted from a prior paper without demonstrated protocol equivalence, and the printed improvement deltas in Table 1 contain arithmetic inconsistencies. The scientific contribution is therefore plausible but not yet established at the level required for acceptance.","major_comments":[{"comment":"The comparison table is not yet sufficient to support the SOTA claim. The text states that results of models other than HARMamba and IF-ConvTransformer are quoted from Miao et al. (2022), but the paper does not establish that Miao et al. used the same data preprocessing, the same 800 ms window with the same overlap, the same user splits, and — for Realdisp — the same 10-user subset. Since the Realdisp section explicitly says only 10 of 17 users are used 'due to incomplete data,' a baseline computed on a different subset would make the reported 4+ point gaps incomparable. The authors should rerun all baselines under the exact protocol used for DecomposeWHAR, or provide a precise per-dataset mapping of the quoted numbers to the exact protocol and user subsets, including evidence that the splits and preprocessing coincide.","section":"Experiments, 'Experimental Results' and Table 1"},{"comment":"The improvement values printed in Table 1 do not consistently match the printed means. For example, the Opportunity macro-F1 improvement for DecomposeWHAR is listed as 6.93%, but the largest visible gap to any runner-up is 72.04 - 67.05 = 4.99 against HARMamba (or 5.91 against DynamicWHAR); the Realdisp accuracy improvement is listed as 4.21%, but 96.64 - 92.58 = 4.06 against DynamicWHAR. Several other deltas are off by 0.1-0.2 points. These values must be recomputed and the reference runner-up for each cell stated explicitly.","section":"Table 1, '(Improvement)' rows"},{"comment":"The text claims that 'Replacing the CSI module with a GCN (GCN as CSI) proposed in DynamicWHAR led to a decline in performance,' but Table 2 contains no 'GCN as CSI' row. This ablation result is therefore not reported and the claim is unsupported. Either add the corresponding row with numbers and standard deviations, or remove the claim from the text.","section":"Ablation Study, Table 2"},{"comment":"For Skoda, the paper describes an 80/10/10 holdout split but does not specify how the split is performed. Because Skoda contains a single subject, a random split without temporal ordering constraints can place adjacent windows from the same activity instance in both training and test sets, inflating accuracy. The authors should describe the split procedure (e.g., contiguous segment boundaries, seed, or leave-one-session-out) and justify that it matches the protocol used for the quoted baselines.","section":"Experimental Settings, Skoda protocol"}],"minor_comments":[{"comment":"There is a typo: 'Readisp' should be 'Realdisp'.","section":"Experimental Settings"},{"comment":"The expression 'T = L S' should be written as T = floor(L/S) or otherwise clarified, since S is the stride and the output length should be an integer.","section":"Equation (1)"},{"comment":"The figure caption says FLOPs are not shown 'due to significant discrepancies,' while the text states that DecomposeWHAR, HARMamba, and DynamicWHAR have FLOPs under 600M. Please reconcile these statements and, if FLOPs are omitted, report them in a table or appendix so the efficiency claim is verifiable.","section":"Figure 3 and Computational Efficiency paragraph"},{"comment":"The '(Improvement)' line after the HARMamba row is unexplained: it is unclear what baseline the HARMamba improvements are computed against and why the values are negative for some metrics. This row should be removed or clearly labeled.","section":"Table 1"},{"comment":"The abbreviation 'MSE' for Modality-Specific Embedding is easy to confuse with Mean Squared Error; a different acronym (e.g., 'MSEmb') would improve readability.","section":"Notation"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid architectural story and the ablations are generally informative, but the evaluation section needs substantial strengthening before the SOTA claim can be accepted. The most serious issue is comparability of quoted baselines, especially on Realdisp where only 10 of 17 users are used. I would encourage the editor to request a full re-benchmarking under a single codebase, or at minimum a detailed protocol-equivalence argument and exact delta recomputation, before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things before you read it. The architecture is a legitimate new composition: per-variable depth-wise convolutions for local temporal features, point-wise fusions at channel and variable level, a Mamba block for global temporal aggregation, and self-attention for cross-sensor interaction. The ablations are consistent with that design story. But the central SOTA claim is not yet solid. The improvements in Table 1 do not match the printed means. On Opportunity, the paper reports +4.02% accuracy and +6.93% macro-F1 for DecomposeWHAR, but the visible gaps to the runner-up are 3.15 and 5.91 respectively. Similar mismatches appear on Realdisp and Skoda. That is a data integrity problem, not a quibble.\n\nThe other big issue is the baseline protocol. Most baselines are quoted from Miao et al. (2022); only HARMamba and IF-ConvTransformer were run by the authors. Realdisp uses 10 of 17 users 'due to incomplete data.' Nothing demonstrates that the quoted baselines used the same user subset, the same sensor placements, or the same windowing. If the protocols differ, the reported 3–5 point margins could be evaluation artifacts.\n\nCredit where due: the decomposition at sensor, variable, and channel levels with reverse-order fusion is a sensible and reasonably novel design in WHAR. The ablation study covers the key components, and the efficiency comparison on a real watch is a nice addition. The paper is clearly written and the equations are checkable.\n\nSofter spots: the text mentions a 'GCN as CSI' ablation that does not appear in Table 2, and the standard deviations are reported without significance tests, so the smaller deltas could be noise. Minor relative to the table issue.\n\nThis paper is for people designing multi-sensor HAR architectures; the decomposition-fusion template is worth studying even if the exact gains need verification. I would accept it for peer review, but the authors must recompute Table 1, fix the delta reporting, and either re-run baselines under identical conditions or provide a rigorous comparability argument. Without that, the SOTA claim should not stand.","headline":"Solid architecture, shaky comparison table: the SOTA claim needs verification before it can be trusted.","tokens_in":12094,"tokens_out":3764,"would_cite":false,"duration_ms":39713,"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 then fusing sensor signals sets a new best for wearable activity recognition on three public datasets.","keywords":["wearable human activity recognition","multi-sensor time series","depth-wise separable convolution","state space model","self-attention","inter-sensor correlation","intra-sensor fusion","activity recognition benchmark"],"falsifier":"Re-run all baseline models on Opportunity, Realdisp, and Skoda under DecomposeWHAR's exact preprocessing, leave-one-user-out splits, and 800 ms windows, and check whether the accuracy and macro-F1 gaps shrink or vanish; the claim fails if DecomposeWHAR no longer leads. For Realdisp, also run the full 17-user protocol.","tokens_in":11146,"feed_emoji":"⌚","tokens_out":8783,"duration_ms":79277,"temperature":0.7,"pith_summary":"This paper argues that multi-sensor wearable activity recognition is hurt by treating all sensor variables with shared convolutional kernels, which blurs the distinction between variables within one sensor and variables across sensors. It proposes DecomposeWHAR, a two-phase network that first decomposes each sensor variable into independent high-dimensional temporal features and then fuses them hierarchically, ending with a Mamba state-space model for global temporal context and a self-attention layer for cross-sensor relationships. The paper reports that this design outperforms existing methods on Opportunity, Realdisp, and Skoda, with accuracy gains of 1.14 to 4.21 points over the runner-up and macro-F1 gains up to 6.93 points. If these results hold, the model is a new best known approach for these benchmarks.","feed_headline":"Sensor-axis decomposition lifts activity recognition by 4.2 points","feed_subtitle":"Separating each sensor channel before fusing it back raises accuracy and F1 on three public wearable datasets.","key_machinery":"The load-bearing mechanism is a two-phase architecture: Modality-Aware Signal Decomposition isolates each sensor, embeds each variable channel separately, and applies depth-wise 1D convolutions so every intra-sensor channel keeps its own temporal kernel; Hierarchical Interaction Fusion then reverses that separation with grouped point-wise convolutions, a Mamba Selective SSM block for global temporal dependencies, and a self-attention layer whose queries, keys, and values are entire per-sensor feature maps, yielding asymmetric inter-sensor correlation scores. This design replaces shared convolutional kernels and symmetric GCN-based sensor fusion, and its ablation shows each stage contributes.","core_discovery":"The central claim is that decomposing multi-sensor signals at sensor, variable, and channel levels, then fusing them in reverse hierarchical order, lets the model capture intra-sensor and inter-sensor spatio-temporal relationships better than shared-kernel convolutions or graph-based sensor fusion. On the three datasets, DecomposeWHAR achieves 78.28% accuracy and 72.04% macro-F1 on Opportunity, 96.64% and 96.10% on Realdisp, and 97.61% and 97.24% on Skoda, beating the runner-up by 1.14 to 4.21 accuracy points and up to 6.93 macro-F1 points. The paper also shows that removing the global temporal aggregation or cross-sensor attention stages sharply degrades performance, which supports the claim that both stages are load-bearing.","pith_inferences":["If the decomposition-fusion pattern is what drives the gains, the same sensor-level, variable-level, and channel-level decomposition could transfer to other multivariate time-series tasks such as EEG analysis or industrial sensor monitoring, but the paper does not test this.","The learned self-attention weights might serve as a diagnostic tool for which body-worn sensors matter most for a given activity, since the attention matrix is asymmetric and inspectable.","A direct re-implementation of all baselines under identical preprocessing would settle whether the reported margins are model gains or protocol effects, especially given the paper's use of 10 of 17 Realdisp users."],"forward_implications":["DecomposeWHAR becomes the new reference point for multi-sensor wearable activity recognition on Opportunity, Realdisp, and Skoda if the reported numbers reproduce.","The depth-wise decomposition eliminates the need for shared convolutional kernels across sensor variables, so per-modality temporal features survive until later fusion.","Inter-sensor relationships no longer depend on a predefined graph; the self-attention module learns them from data and captures asymmetry between sensors.","The model retains acceptable computational cost, with FLOPs under 600M and competitive inference time on a wearable watch, so the accuracy gains are not purchased with disproportionate compute.","Ablations show the cross-sensor attention and global temporal aggregation stages carry most of the improvement; removing both degrades macro-F1 by 9.27 to 19.51 points across datasets."],"supporting_citations":[{"why":"Source of the DynamicWHAR baseline and most quoted comparison numbers; its GCN-based inter-sensor fusion is the main alternative architecture being challenged.","marker":"(Miao et al. 2022)"},{"why":"Supplies the Mamba Selective SSM used by the Global Temporal Aggregation module for long-range dependencies.","marker":"(Dao and Gu 2024)"},{"why":"Introduces depthwise separable convolution, which the decomposition phase adapts to 1D temporal feature extraction per sensor channel.","marker":"(Chollet 2017)"},{"why":"Provides the Opportunity benchmark used in the main evaluation.","marker":"(Roggen et al. 2010)"},{"why":"Provides the Realdisp benchmark, with the paper using 10 of its 17 users.","marker":"(Baños et al. 2012)"},{"why":"Provides the Skoda benchmark for the single-user car-maintenance evaluation.","marker":"(Stiefmeier et al. 2008)"},{"why":"Defines the Attend and Discriminate baseline and the Skoda 80/10/10 hold-out protocol.","marker":"(Abedin et al. 2021)"},{"why":"Defines DeepConvLSTM, the shared-convolution baseline this work argues is suboptimal.","marker":"(Ordóñez and Roggen 2016)"},{"why":"Provides the HARMamba Mamba-based baseline, which the paper implements itself for comparison.","marker":"(Li et al. 2024)"}],"fun_headline_variants":["Decompose-then-fuse sensor signals boosts HAR accuracy","Multi-sensor HAR improved by splitting then fusing channels","Splitting sensor axes before fusion lifts activity recognition","DecomposeWHAR: decompose and fuse for smarter wearables","Intra- and inter-sensor fusion breaks HAR accuracy records"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume the quoted baseline numbers were produced under the same preprocessing, user splits, and window size as DecomposeWHAR; if the protocols differ, especially on Realdisp where only 10 of 17 users are used, the reported margin may partly reflect evaluation differences rather than model quality.","fun_headline_variants_meta":{"raw":{"variants":["Decompose-then-fuse sensor signals boosts HAR accuracy","Multi-sensor HAR improved by splitting then fusing channels","Splitting sensor axes before fusion lifts activity recognition","DecomposeWHAR: decompose and fuse for smarter wearables","Intra- and inter-sensor fusion breaks HAR accuracy records"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1384,"prompt_tokens":928,"completion_tokens":456,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":372}},"tokens_in":544,"tokens_out":456,"duration_ms":4618,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:49:27.168841+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run all baseline models on Opportunity, Realdisp, and Skoda under DecomposeWHAR's exact preprocessing, leave-one-user-out splits, and 800 ms windows, and check whether the accuracy and macro-F1 gaps shrink or vanish; the claim fails if DecomposeWHAR no longer leads. For Realdisp, also run the full 17-user protocol.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the DynamicWHAR baseline and most quoted comparison numbers; its GCN-based inter-sensor fusion is the main alternative architecture being challenged."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Mamba Selective SSM used by the Global Temporal Aggregation module for long-range dependencies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces depthwise separable convolution, which the decomposition phase adapts to 1D temporal feature extraction per sensor channel."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Opportunity benchmark used in the main evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Skoda benchmark for the single-user car-maintenance evaluation."}],"review_version":1}