{"id":"09f3f8a6-9b26-43fc-b018-c84d8b270a0a","arxiv_id":"2507.03250","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A subject-reweighted contrastive loss improves cross-subject generalization for human activity recognition across unimodal, multimodal, and supervised settings.","lead":"SICL reweights same-subject negatives in a contrastive loss to push human activity recognition models away from subject-specific cues. On three benchmarks it reports consistent top-1 accuracy gains over standard contrastive methods, with the largest relative gain around 12 percent on DARai foot pressure.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3)'s QSi is never defined as a formula; the entire SICL result rests on an ambiguous prose description, so the reported gains cannot be attributed to the stated re-weighting without an explicit definition.","rationale":"I read the preprint in good faith: the central idea is coherent, the experiments span three datasets and several contrastive frameworks, and the reported improvements are directionally consistent. However, the exact form of QSi is the single most load-bearing detail in the method. Equation (3) defines the SICL denominator in terms of QSi, but the paper supplies only a loose prose description of how QSi is computed. Without a closed-form definition, the method cannot be reproduced from the paper alone, and the claimed gains cannot be confidently attributed to the stated mechanism rather than to an incidental choice in the code. This is not an attack on the authors: a public code repository is linked, so the ambiguity is checkable. The concrete test I propose extracts the actual formula from the repository and compares it against the two most natural readings of the prose, using the UTD-MHAD inertial row where the SimCLR-to-SICL gap is largest among the unimodal results. If the test shows that plausible variants of QSi change accuracy by more than the reported gain, then the central claim is conditional until the definition is supplied and the implementation is validated. This matches the reader's weakest assumption, so I do not change the CONDITIONAL verdict; the concern reinforces the condition rather than overturning the paper.","tokens_in":972,"tokens_out":1107,"duration_ms":59060,"concrete_test":"Run the released GitHub code on UTD-MHAD inertial, print the computed QSi values, and state the exact formula used in the repository. Then re-run the same experiment with QSi replaced by a constant 1 (standard SimCLR denominator) and by a softmax-normalized interpretation of the Section 4.1 prose. If the top-1 accuracy moves by more than the reported SICL gain over SimCLR (0.0218 absolute on UTD-MHAD inertial), the claimed improvement is not pinned down by the paper's text and depends on the hidden implementation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing element of SICL is the per-sample weighting QSi that multiplies the same-subject negative sum in Eq. (3), and its supervised analogue QSij in Eq. (5). Nowhere in Section 4.1 is QSi written as an equation. The only specification is the prose: 'normalizing the exponentiated cosine similarities of same-subject negatives, summing these values, and then dividing by their average.' This admits multiple implementations: normalization could be by sum (softmax), by max, or by L2 norm; the values could be normalized per anchor or globally per batch; and 'dividing by their average' could produce a scalar that multiplies the whole same-subject sum or per-negative weights under a different normalization. Because D_i in Eq. (3) is the entire mechanism, these choices change which pairs are pushed apart and by how much, and can even alter the intended sign or magnitude of the same-subject penalty. The reported gains (e.g., UTD-MHAD inertial 0.7173 to 0.7391 for SimCLR+SICL) are small enough that plausible alternative readings of the prose could produce comparable or contradictory differences. A secondary issue is that the paper never states whether two different time windows from the same subject performing the same activity are treated as negatives; if they are, re-weighting them apart may suppress activity structure as well as subject structure, and no experiment separates these two effects.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Subject-Invariant Contrastive Learning (SICL), a modification of the contrastive loss that re-weights negative pairs drawn from the same subject. The goal is to reduce subject-specific information in learned embeddings and improve generalization to unseen subjects in human activity recognition. The loss is applied to unimodal and multimodal contrastive frameworks, as well as supervised contrastive learning. The authors evaluate on UTD-MHAD, MMAct, and DARai, reporting consistent improvements over baselines in linear evaluation, fine-tuning, and supervised settings.","tokens_in":9646,"tokens_out":7208,"duration_ms":74233,"significance":"The paper addresses an important and practical problem: subject variability in HAR causes domain shift, and standard contrastive learning does not account for it. The proposed loss is conceptually simple and can be plugged into several existing methods (SimCLR, Barlow Twins, VICReg, CMC, CMC-CMKM, SupCon). The authors provide a public code repository and use openly available datasets, and they include a motivating analysis (ANOVA, cosine similarity distributions) showing that subject shift exists. If the mechanism is made precise and the empirical results are verified with proper statistical support, the method would be a useful contribution to the HAR community. However, at present the key weighting function is not formally defined and the experimental evaluation lacks error bars, which tempers the strength of the empirical claims.","major_comments":[{"comment":"The weighting function QSi, which is the central new mechanism of SICL, is never defined as a formula. The prose description in Section 4.1 ('normalizing the exponentiated cosine similarities of same-subject negatives, summing these values, and then dividing by their average') admits multiple implementations: normalization could be by sum, by max, or by L2 norm; it could be computed per anchor or globally per batch; and the 'average' division could produce a scalar multiplier or a set of per-pair weights. Because QSi (and its supervised analogue QSij in Eqs. (5) and (7)) multiplies the entire same-subject negative sum in Eq. (3), these choices change the optimization landscape and could even alter the sign or relative magnitude of the penalty. The reported gains in Tables 1-4 are therefore not attributable to a specific re-weighting scheme. Please provide the exact functional form of QSi and QSij, or a reference to a specific equation in [17] or [18] that defines them.","section":"Section 4.1, Eq. (3)"},{"comment":"The experimental protocol for the unseen-subject evaluation is underspecified. The paper does not state how subjects are partitioned into training and test sets for each dataset (e.g., the number of training subjects, whether multiple splits are used, or whether results are averaged across seeds). This is critical because some reported improvements are small: for example, in Table 1 on DARai inertial data, SimCLR+SICL (0.1610) and SICL (0.1615) differ by only 0.0005, and the absolute gain over SimCLR (0.1498) is about 0.011. Without error bars or repeated runs, the claim of 'consistent improvements' across all cells is not statistically supported. Please provide variance estimates (e.g., standard deviations over at least three seeds) and a precise description of the subject split.","section":"Section 5.1, Tables 1-4"},{"comment":"In the self-supervised contrastive setup of Eq. (1), the negative set A(i) contains all other instances in the batch, including time windows from the same subject performing the same activity as the anchor. Reweighting all same-subject negatives via QSi in Eq. (3) therefore also pushes apart representations of identical activities within a subject. The paper does not analyze this effect, nor does it provide a control experiment that differentiates between suppressing subject identity and suppressing same-activity structure. Figure 3 only shows cosine similarities of all pairs versus intra-subject pairs; it does not condition on activity labels. Since the method's stated goal is to remove subject-specific cues while retaining activity information, an experiment that varies the composition of S(i) (e.g., excluding same-activity negatives) would be necessary to support that causal claim.","section":"Section 4.1, Eq. (1)-(3)"},{"comment":"The paper claims that SICL learns subject-invariant representations, but no direct measure of subject invariance is reported. The only evidence is downstream activity classification accuracy and the aggregate cosine similarity distributions in Figure 3. A more direct test would be to train a classifier to predict subject identity from the learned embeddings: if the method indeed suppresses subject-specific cues, subject-classification accuracy should drop substantially relative to SimCLR while activity accuracy is maintained. This experiment would strengthen the central claim and is a natural complement to the ANOVA analysis in Section 3.","section":"Section 5.3.1, Figure 3"}],"minor_comments":[{"comment":"The caption of Figure 2 is incomplete: it ends with 'In the seen subject setting.' followed by nothing, and the sentence about the t-SNE plot is cut off. Please fix the caption.","section":"Figure 2 caption"},{"comment":"The sentence 'This framework has been adapted for sensor-based human activity recognition by employing signal augmentations to form positive pairs during contrastive loss computation [9]. Negative samples are again drawn uniformly from the data. Recent methods have suggested more informative negative sampling strategies... However, these methods do not consider distribution shifts in HAR data during the sampling of negatives.' is missing a verb or subject in the first clause; the phrase 'without considering their informativeness for learning discriminative representations' in the previous sentence is unclear. Consider rewriting for clarity.","section":"Introduction, paragraph 3"},{"comment":"For DARai, the paper states 'We use a subset of the dataset consisting of 20 subjects' but does not specify which subjects or how the subset is selected. Please provide this information to ensure reproducibility.","section":"Section 5.1"},{"comment":"The sentence 'We use the same temperature parameters as in [11]' is vague. Please list the temperature values used for each dataset and modality, or provide a reference to the specific configuration.","section":"Section 5.2"},{"comment":"The abstract claims improvement 'up to 11%', but the largest relative improvement visible in Table 1 is approximately 10% (Barlow Twins+SICL vs. Barlow Twins on DARai inertial). Please reconcile this number or clarify the calculation.","section":"Abstract and Section 6"}],"recommendation":"major_revision","confidential_remarks":"The paper relies on the authors' prior work [17] and [18] for the weighting scheme, but it is not self-contained in this manuscript. The undefined QSi is the main blocking issue; it is fixable, so I recommend major revision rather than rejection. The lack of error bars and the small absolute gains on DARai are also concerns, but these can be addressed with additional experiments and reporting. This is a workshop-length paper, so the expectations for statistical rigor may be lower, but the central claim depends on a precise definition of the loss."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a legitimate, small extension of the authors' prior taxonomic reweighting trick, applied to subject identity in sensor-based HAR. The empirical pattern is consistent across three datasets and several SSL methods, so I don't think it's a fluke. But the paper has one load-bearing ambiguity: the weighting function QSi in Eq. (3) is only described in prose, never written as a formula, and the entire mechanism hangs on it.\n\nWhat's actually new? Nothing conceptual beyond substituting subject labels for taxonomy in [17,18]. That's fine as a contribution—it works for HAR, and the authors are honest about the lineage. The experiments are reasonably thorough: unimodal, multimodal, supervised, and a pre-training/fine-tuning variant. I appreciate the attempt to show subject shift with ANOVA and the cosine similarity plots; those motivate the loss.\n\nThe soft spots, in order of importance. First, QSi. The paper says it is calculated 'by normalizing the exponentiated cosine similarities of same-subject negatives, summing these values, and then dividing by their average.' That can be read at least three ways—softmax, divide-by-max, L2 norm—and the resulting weighting would change which pairs get pushed apart and by how much. Because the reported gains are small (a couple of absolute points), alternative readings of the prose could plausibly produce those differences for reasons unrelated to the stated intent. I'd want the exact formula, plus a sentence about whether same-subject, same-activity windows are ever negatives, and if so, why the loss doesn't suppress activity structure. Second, no error bars or repeated seeds. For a method that claims universal improvement, I'd like to know the variance. Third, the DARai evaluation uses a 20-subject subset, which is fine, but the selection criteria are not justified. Fourth, the self-reference is heavy—[17,18] supply the structure—but that's not a flaw per se, given the lineage is explicit.\n\nIs the central argument sound? I think the general direction—reweighting intra-subject negatives to push out subject-specific cues—is plausible and the results support it, at least in aggregate. The ambiguity of QSi is a fixable clarity issue, not a fundamental contradiction.\n\nWho's this for? Someone working on sensor-based HAR who wants a plug-in loss to try on their own data. It's not a big theoretical advance, but it's a practical, testable idea.\n\nRecommendation: I'd send it to peer review, but with the explicit equation, error bars, and the same-activity question as required revisions. It's below the bar for a top venue as-is, but it's a legitimate workshop-to-conference level contribution. I'd cite it if I were doing follow-up on subject-invariant contrastive losses.","headline":"Modest but plausible extension of the authors' own taxonomy reweighting to subject labels; consistent gains across three datasets, but the load-bearing weighting term QSi is never defined as an equation, which needs fixing before the result is fully trustworthy.","tokens_in":10142,"tokens_out":2041,"would_cite":true,"duration_ms":23891,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that contrastive HAR training embeds subject-specific cues that hurt generalization, and that reweighting same-subject negatives in the loss suppresses those cues and improves accuracy on unseen subjects.","keywords":["subject-invariant contrastive learning","human activity recognition","contrastive loss","distribution shift","self-supervised learning","multimodal learning","unseen-subject generalization","sensor data"],"falsifier":"Run SICL and SimCLR on the same subject-disjoint splits, then rerun SICL with the same-subject weight set to 1 everywhere so the loss reduces to the standard contrastive denominator; if held-out-subject accuracy does not change, the reported gains do not come from the subject weighting. A complementary check is to use data where each subject performs only one activity, so subject identity and activity are perfectly confounded; there, suppressing subject cues should hurt SICL relative to SimCLR.","tokens_in":9152,"feed_emoji":"📈","tokens_out":9120,"duration_ms":88455,"temperature":0.7,"pith_summary":"Human-activity-recognition models trained with contrastive learning tend to encode who is moving rather than what activity is being performed, because the same activity looks different across subjects. This paper introduces Subject-Invariant Contrastive Learning (SICL), an extra term for the contrastive loss that separates negative pairs drawn from the same subject, forcing the encoder to attend to activity-specific cues. SICL requires only subject labels, not activity labels, and it can be layered on top of existing self-supervised losses. On three public benchmarks the authors report consistent top-1 accuracy gains on held-out subjects, up to 11 percent over standard contrastive learning, and show the same reweighting improves supervised contrastive learning and multimodal setups.","feed_headline":"Reweight same-subject negatives to boost activity recognition","feed_subtitle":"Adding one reweighting term to the contrastive loss suppresses subject cues and improves accuracy on held-out subjects.","key_machinery":"The load-bearing object is the SICL loss, an altered contrastive denominator. With $z_i$ the anchor embedding and $\\tau$ the temperature, the loss is $L_{\\mathrm{SICL}} = -\\sum_i \\log\\left( \\frac{\\exp(z_i \\cdot z_j / \\tau)}{D_i} \\right)$, where $D_i = Q_{S_i} \\sum_{s \\in S(i)} \\exp(z_i \\cdot z_s / \\tau) + \\sum_{k \\notin S(i)} \\exp(z_i \\cdot z_k / \\tau)$. Here $S(i)$ is the set of negatives sharing the anchor's subject; $Q_{S_i}$ is computed batch-wise by exponentiating cosine similarities of same-subject negatives, summing them, and dividing by their average. Multiplying the same-subject sum by $Q_{S_i}$ inflates that part of the denominator, so minimizing the loss pushes anchors away from other samples of the same subject. Subject labels are used only during training, so the learned encoder remains usable without labels at test time.","core_discovery":"The paper's central claim is that the distribution shift induced by subject variability is a primary reason contrastive HAR representations fail on unseen subjects. The authors first show evidence for the shift: on the DARai dataset, spectral entropy and dominant frequency differ significantly across subjects in one-way ANOVAs, seen-subject accuracy exceeds unseen-subject accuracy for SimCLR, and intra-subject pairs have higher cosine similarity than all pairs after standard contrastive training. SICL then changes the contrastive denominator by splitting negatives into same-subject and other-subject sets and multiplying the same-subject terms by a batch-wise weight $Q_{S_i}$ built from exponentiated cosine similarities of same-subject negatives. This raises the penalty for leaving embeddings of the same person close together, so the encoder must separate samples by activity rather than by subject identity. Across UTD-MHAD, MMAct, and DARai, the paper reports consistent top-1 accuracy gains over SimCLR, hard-negative contrastive loss, Barlow Twins, and VICReg, and also over the CMC and CMC-CMKM multimodal frameworks and the supervised contrastive loss.","pith_inferences":["The same reweighting recipe should transfer to other sensor domains with a known recording source, such as per-patient medical time series or per-device wearables, where source identity creates a similar distribution shift.","Because SICL needs only subject labels, it could be applied in federated or privacy-preserving settings where each client is one subject and activity labels are unavailable.","A closed-form definition of $Q_{S_i}$ would let practitioners test sensitivity to the weighting scheme, and making this weight a learned function of similarity could adapt suppression strength per subject.","SICL's gain could serve as a diagnostic: if adding the term does not help on a given dataset, the dominant distribution shift in that dataset is probably not subject identity."],"forward_implications":["Adding SICL to SimCLR, Barlow Twins, or VICReg raises linear-evaluation top-1 accuracy on held-out subjects across UTD-MHAD, MMAct, and DARai.","SICL extends to multimodal contrastive frameworks: it improves CMC and CMC-CMKM on inertial-plus-skeleton and inertial-plus-foot-pressure inputs.","Using SICL as pretraining before cross-entropy fine-tuning beats no pretraining, SimCLR pretraining, and hard-negative pretraining in the reported tables.","Applying subject reweighting to the supervised contrastive loss (SI-SupCon) improves over both SupCon and SupHCL, so the mechanism transfers to label-aware training.","Trained representations become more subject-invariant: the cosine-similarity distribution of intra-subject pairs moves closer to the all-pairs distribution."],"supporting_citations":[{"why":"The DARai dataset supplies the real-world multimodal daily-activity data used for the subject-variability ANOVA and for the main evaluations.","marker":"[4]"},{"why":"SimCLR is the base contrastive learning method whose uniform negative sampling SICL modifies, and the main baseline for the seen-versus-unseen subject comparison.","marker":"[5]"},{"why":"Provides the sensor-HAR contrastive learning setup, including the 1D-CNN plus transformer encoder and signal augmentations, used in all experiments.","marker":"[9]"},{"why":"Supplies the batch-wise negative-reweighting strategy that SICL adapts from taxonomic hierarchy to subject labels.","marker":"[17]"},{"why":"The supervised contrastive loss that SICL extends into SI-SupCon.","marker":"[19]"},{"why":"Contrastive Multiview Coding is the base multimodal framework that SICL modifies into SI-CMC.","marker":"[20]"},{"why":"CMC-CMKM is the cross-modal knowledge-mining baseline that SICL also modifies and compares against.","marker":"[21]"},{"why":"Hard negative contrastive loss (HCL) is one of the main baseline losses compared in the unimodal and pretraining tables.","marker":"[25]"}],"fun_headline_variants":["Reweight same-subject negatives to boost activity recognition","Suppress subject-specific cues to improve HAR generalization","SICL: reweight negatives by subject to boost activity recognition","One reweighting term boosts activity recognition across subjects","Subject-invariant contrastive loss improves HAR accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that reweighting negative pairs from the same subject pushes those embeddings apart without harming the separation between different activities, and that subject identity is the dominant shift in the data.","fun_headline_variants_meta":{"raw":{"variants":["Reweight same-subject negatives to boost activity recognition","Suppress subject-specific cues to improve HAR generalization","SICL: reweight negatives by subject to boost activity recognition","One reweighting term boosts activity recognition across subjects","Subject-invariant contrastive loss improves HAR accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000233,"raw_usage":{"total_tokens":1503,"prompt_tokens":965,"completion_tokens":538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":461}},"tokens_in":581,"tokens_out":538,"duration_ms":5695,"temperature":1.0,"reasoning_tokens":461,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:14:44.085753+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SICL and SimCLR on the same subject-disjoint splits, then rerun SICL with the same-subject weight set to 1 everywhere so the loss reduces to the standard contrastive denominator; if held-out-subject accuracy does not change, the reported gains do not come from the subject weighting. A complementary check is to use data where each subject performs only one activity, so subject identity and activity are perfectly confounded; there, suppressing subject cues should hurt SICL relative to SimCLR.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The DARai dataset supplies the real-world multimodal daily-activity data used for the subject-variability ANOVA and for the main evaluations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SimCLR is the base contrastive learning method whose uniform negative sampling SICL modifies, and the main baseline for the seen-versus-unseen subject comparison."},{"cited_title":"Datasets We evaluate the performance of our proposed approaches on three benchmark multimodal datasets: UTD-MHAD [22], MMAct [23], and DARai [4]","cited_arxiv_id":null,"evidence_quote":"Provides the sensor-HAR contrastive learning setup, including the 1D-CNN plus transformer encoder and signal augmentations, used in all experiments."},{"cited_title":"Contrastive self-supervised learning for sensor-based human activity recognition,","cited_arxiv_id":null,"evidence_quote":"The supervised contrastive loss that SICL extends into SI-SupCon."},{"cited_title":"Multimodal Contrastive Learning with Hard Negative Sampling for Human Activity Recognition","cited_arxiv_id":"2309.01262","evidence_quote":"Contrastive Multiview Coding is the base multimodal framework that SICL modifies into SI-CMC."},{"cited_title":"Contrastive learning with cross-modal knowledge mining for multimodal human activity recognition,","cited_arxiv_id":null,"evidence_quote":"CMC-CMKM is the cross-modal knowledge-mining baseline that SICL also modifies and compares against."},{"cited_title":"Diversify: A gen- eral framework for time series out-of-distribution detection and generalization,","cited_arxiv_id":null,"evidence_quote":"Hard negative contrastive loss (HCL) is one of the main baseline losses compared in the unimodal and pretraining tables."}],"review_version":1}