{"id":"376d5132-fe7b-4f24-ab50-96f733c6a5ed","arxiv_id":"2412.13594","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new regularization loss on an element-wise product of features and classifier weights improves domain-generalized human activity recognition across four benchmarks.","lead":"This paper proposes CCIL, a training regularizer for sensor-based activity recognition that pulls an element-wise feature-weight 'concept matrix' of same-class samples together to improve cross-person and cross-dataset generalization. It reports consistent accuracy gains over domain generalization baselines on four public HAR datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported margins over the strongest baseline are small (1.3-2.1 pts) in three of four settings, no error bars are given, and Eq. (5) is internally inconsistent with the ablation text, so the 'substantially outperforms' claim is not yet established.","rationale":"The reader's weakest assumption concerns instability or degeneracy of the momentum-updated class-mean target. I do not think that is the most load-bearing issue: the paper's own W/lambda=1 ablation, if interpreted as a frozen target, still improves over ERM by about 9 points, so a stable target is not essential to the method's success. The more decisive issue is whether the reported advantages are real rather than selection noise, and whether the algorithm is specified precisely enough to reproduce. The small margins in Tables 2-4, the absence of error bars, and the ambiguity of the three-trial selection protocol directly threaten the 'substantially outperforms' claim. The Eq. (5) coefficient inconsistency is an additional internal error: as printed, lambda=1 is a current-batch target and lambda=0 is frozen, but the ablation text and Table 5 correspond to the opposite convention. This ambiguity affects every reported result because the momentum update is the core mechanism. The method is plausible and Table 1 shows large margins, so rejection is not warranted, but the current evidence supports only a conditional acceptance pending code release, more seeds, and a corrected Eq. (5).","tokens_in":15073,"tokens_out":13054,"duration_ms":126559,"concrete_test":"Obtain the official code or reimplement from a corrected Eq. (5), then run CCIL and DIVERSIFY on all four settings with 10 independent seeds under the same source-domain validation protocol, recording test accuracy per seed. Compute the paired 95% bootstrap confidence interval for the mean difference (CCIL minus DIVERSIFY) on each setting. If the interval includes zero for cross-position, cross-dataset, or one-person-to-another, the claim of substantial superiority is not supported; also verify which lambda convention reproduces Table 5.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: CCIL substantially outperforms state-of-the-art DG methods on four HAR settings. The evidence is three-trial point estimates with no standard deviations, and in three of the four settings the advantage over the strongest baseline DIVERSIFY is only 1.3, 1.5, and 2.1 accuracy points. On PAMAP2 in the cross-dataset setting, CCIL actually trails DIVERSIFY (59.6 vs 59.9). Without error bars, a paired analysis, or a stated number of seeds, these gaps are within typical run-to-run noise for HAR. The sentence 'All methods were adjusted to report the average best performance over three trials' does not clarify whether the test set was used for model selection, which would inflate accuracy. Reproducibility is further compromised by an internal inconsistency in the core update: Eq. (5) defines the momentum update so that lambda=1 yields a current-batch target and lambda=0 yields a frozen target, while the ablation text says the opposite. Table 5's W/lambda=0 and W/lambda=1 numbers match the standard EMA convention, suggesting the printed equation is reversed. As a result, the exact method behind Tables 1-4 cannot be reconstructed from the paper alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CCIL, a regularization method for sensor-based human activity recognition under domain shift. For each sample, it forms a D x C 'concept matrix' M = W * z from the feature vector z and classifier weights W, and adds a loss term L_CMS that penalizes the squared distance between each sample's concept matrix and a class-mean target M_hat_c, which is updated by momentum. The total objective is L = L_CE + alpha * L_CMS. The method is evaluated on DSADS, PAMAP2, USC-HAD, and UCI-HAR under cross-person, cross-position, cross-dataset, and one-person-to-another settings, and compared against ERM, DANN, CORAL, Mixup, GroupDRO, RSC, ANDMask, GILE, AdaRNN, and DIVERSIFY.","tokens_in":15365,"tokens_out":5545,"duration_ms":50352,"significance":"If the reported gains are statistically robust, CCIL would be a simple, model-agnostic regularizer with a clear intuition that goes beyond feature-invariance by also accounting for classifier weights. The paper has several strengths: coverage of four datasets and four shift settings, a source-domain validation protocol, an ablation study, a parameter-sensitivity analysis, a Transformer extensibility experiment, and pseudo-code. However, the central claim of consistent and substantial improvement over state-of-the-art methods is not yet established because the reported margins are small in several settings, no error bars are given, and the description of the momentum update is internally inconsistent. These issues are load-bearing for the paper's empirical contribution.","major_comments":[{"comment":"The momentum update in Eq. (5) is written as M_hat^t = (1 - lambda) * M_hat^{t-1} + lambda * batch_mean, which means lambda=1 replaces the target with the current batch and lambda=0 freezes the initial target. The ablation text states the opposite: 'W/lambda = 0' is described as using the mean value dynamically calculated from the current batch, and 'W/lambda = 1' is described as keeping M_hat fixed from the initial pretrained model. The numbers in Table 5 (W/lambda=0: 85.5, W/lambda=1: 89.3) and the sensitivity analysis in Figure 4 are consistent with the standard EMA convention in Eq. (5), not with the ablation text. Only one convention can be the one that produced Tables 1-4; please correct Eq. (5) or the text and explicitly state the exact update used.","section":"Methodology, Eq. (5) vs. Ablation Study and Table 5"},{"comment":"The paper states that 'All methods were adjusted to report the average best performance over three trials' and that hyperparameters are tuned per method and per dataset, but no standard deviations, per-seed results, or model-selection criterion are reported. This matters because the margins over the strongest baseline are small in three of the four settings: 1.3 points in Table 2 (cross-position DSADS), 1.5 points in Table 3 (cross-dataset average), and 2.1 points in Table 4 (one-person-to-another). With three trials and per-method hyperparameter tuning, these gaps are within typical run-to-run variation for HAR models. Please report the number of seeds, standard deviations, the exact rule for selecting the reported checkpoint (source validation vs. test-set selection), and ideally a paired statistical comparison.","section":"Implementation Details and Tables 1-4"},{"comment":"The claim that CCIL 'consistently' beats the state of the art is not supported on every target: in the cross-dataset setting, target 3 (PAMAP2) shows CCIL at 59.6% versus DIVERSIFY at 59.9%. The average is higher (52.6 vs. 51.1), but the per-target result contradicts the unqualified wording. Please qualify the claim to say CCIL achieves the best average accuracy while not winning every target domain.","section":"Table 3 and Experimental Results"},{"comment":"The protocol description says four pairs of subjects are used - (0,1), (2,3), (4,5), and (6,7) - and that the model generalizes 'from the second subject in each pair to the first.' However, Table 4 reports only three target columns, labeled 0, 1, and 2 as DSADS, USC-HAD, and PAMAP2. It is unclear whether the reported accuracy is averaged over all four pairs for each dataset, or whether only one pair per dataset is used, and how the pair-to-target mapping is defined. Please specify the exact data split and the aggregation rule over the four pairs.","section":"Experimental Setup, One-person-to-another setting"}],"minor_comments":[{"comment":"There are several typographical errors, including 'vanialla' in the Methodology section, 'popuar' in the Problem Formulation section, 'model-agonistic' in the Experimental Results section, and 'E MR' in the Comparative Methods subsection. These should be corrected.","section":"Throughout"},{"comment":"The notation '|yi = c|' in Eq. (5) is undefined and is described as 'the sample corresponding to the c-th class.' It should be defined precisely as the number of samples in the batch whose label is c, and the denominator should be stated as a count.","section":"Methodology, Eq. (5)"},{"comment":"In Step 6, the pseudo-code says 'Using output features z and classifier weight W to calculate the mean values M_hat,' but M_hat is an exponential moving average, not a direct batch computation. This step should be rewritten to match the corrected Eq. (5).","section":"Algorithm 1"},{"comment":"The terms 'W/lambda = 0' and 'W/lambda = 1' are described in a way that is opposite to Eq. (5). After fixing the equation, the ablation text and the definition of 'initial pretrained model' need to be made consistent; otherwise the reader cannot determine whether pretraining was actually used.","section":"Ablation Study"},{"comment":"The axes in Figure 1 are labeled '1' and '2' without explanation. Please label the axes as subject index and sensor position, or add a caption that defines these axes, so that the illustration of domain shift is clear.","section":"Figure 1"},{"comment":"In Table 6, the column 'Subject' lists the number of subjects in each dataset; renaming it to 'Number of subjects' would avoid ambiguity, and the 'Sampling rate' column should state the unit (Hz) directly in the header or in the caption.","section":"Table 6"}],"recommendation":"major_revision","confidential_remarks":"The idea is simple and potentially useful, and the benchmark coverage is broad. The main obstacle is empirical verification: the paper currently reports three-trial point estimates with no error bars, and the internal inconsistency between Eq. (5) and the ablation text prevents the method from being reconstructed exactly. These issues are fixable, but they are central to the paper's contribution, so I cannot recommend acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is the concept matrix: the element-wise product of penultimate features and classifier weights, regularized so same-class samples have similar matrices. That is a legitimate new object, and the paper is honest that it sits between feature-invariance and logit-invariance. The method is simple, cheap, and clearly described: just add the CMS loss to cross-entropy. I buy the motivation, and the ablation showing the concept matrix beating both feature-only and logit-only variants on DSADS is the most convincing part of the paper.\n\nWhat the paper does well: four shift settings, four public datasets, a reasonable baseline set that includes DIVERSIFY, and a Transformer extensibility check. The central claim is empirical, and the coverage is genuinely broad for this subfield. The t-SNE figures are suggestive rather than probative, but they are consistent with the story.\n\nThe soft spots are real but not disqualifying. In three of four settings the margin over DIVERSIFY is 1.3, 1.5, and 2.1 points, with no standard deviations and only 'average best performance over three trials.' On PAMAP2 in cross-dataset, CCIL actually trails DIVERSIFY. Given typical HAR noise, those gaps are not yet 'substantially outperforms.' The per-method hyperparameter tuning and no code release make it harder to check, though the sensitivity analysis shows CCIL is robust across alpha and lambda.\n\nThe stress-test note about Eq. (5) is correct and worth flagging: the equation as printed has lambda=1 giving a current-batch target and lambda=0 giving a frozen target, while the ablation text says the opposite. The Table 5 numbers (W/lambda=0 at 85.5, W/lambda=1 at 89.3) match the standard EMA convention, so the printed equation appears reversed. That is an internal inconsistency that needs fixing, but it is fixable and does not sink the method.\n\nThe circularity concern is minor: the regularizer is a training objective, not a fitted claim. The weakest assumption is real though: if features or classifier weights collapse, the CMS loss can be minimized without improving generalization, and nothing in the objective prevents that degeneracy. The authors do not discuss this failure mode.\n\nWho this is for: anyone working on sensor-based HAR domain generalization. It deserves a serious referee. I would accept it with requests for error bars or more seeds, a corrected Eq. (5), a clear statement of model selection, and ideally code. The core idea is worth engaging with even though the headline margin claims need tempering.","headline":"A simple, plausible regularizer for domain-generalized HAR with real benchmark coverage, but the empirical margins are thin and the paper's own momentum equation contradicts its ablation text.","tokens_in":15845,"tokens_out":637,"would_cite":true,"duration_ms":15392,"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":"This paper argues that aligning each sample's concept matrix—the per-feature, per-class contributions of features times classifier weights—with its class mean makes sensor-based activity recognition generalize across unseen people…","keywords":["human activity recognition","domain generalization","concept matrix","sensor-based activity recognition","cross-person generalization","momentum update","logit invariance","wearable sensors"],"falsifier":"Train CCIL while tracking the rank and norm of the classifier weights $W$ and the per-class feature covariance; if $L_{\\mathrm{CMS}}$ decreases mainly through shrinking feature or weight norms, or through a low-rank collapse of $W$, the claimed alignment mechanism is not what drives the accuracy gains. A direct check is to re-run the experiments with a norm-preserving reparameterization that keeps $\\|W\\|$ and $\\|z\\|$ approximately fixed and see whether the gains survive.","tokens_in":14899,"feed_emoji":"⌚","tokens_out":11032,"duration_ms":87824,"temperature":0.7,"pith_summary":"Sensor-based activity-recognition models often fail when tested on people, body positions, or datasets they have not seen during training. This paper tries to establish that a single regularization term can close much of that gap. The term enforces a new kind of invariance: for every sample, the concept matrix—the element-wise products of the feature values with the classifier weights that form each logit contribution—should sit close to the mean concept matrix of its activity class. Added to standard cross-entropy, this loss is reported to beat existing domain-generalization methods on four public HAR benchmarks under cross-person, cross-position, cross-dataset, and one-person-to-another settings. The practical stake is that a model trained on available subjects could then be deployed on new users, including groups whose data cannot be collected in advance.","feed_headline":"New loss lifts cross-person activity recognition up to 10 points","feed_subtitle":"Aligning each sample's feature-weight contributions to its class mean also helps unseen positions and datasets.","key_machinery":"The central object is the concept matrix $M_i$, defined entrywise as $(M_i)_{j,c} = W_{j,c} z_j$: each entry is one feature's contribution to one class logit before summation. The argument is carried by two pieces working together: the loss $L_{\\mathrm{CMS}}$, which penalizes within-class scatter of these matrices around their class mean, and the momentum update $\\hat{M}_c^t = (1-\\lambda)\\hat{M}_c^{t-1} + \\lambda \\bar{M}_c$, where $\\bar{M}_c$ is the mean concept matrix for class $c$ in the current batch. Together they provide a fine-grained, classifier-aware alignment that the paper's ablations show is stronger than aligning features alone or logits alone.","core_discovery":"The paper's central claim is that a well-generalized activity model should produce similar concept matrices for samples of the same activity, regardless of domain. For input sample $i$, the concept matrix is $(M_i)_{j,c} = W_{j,c} z_j$, the contribution of feature dimension $j$ to class logit $c$. CCIL adds the categorical concept invariance loss $L_{\\mathrm{CMS}} = \\frac{1}{N_b} \\sum_c \\sum_{\\{i:y_i=c\\}} \\|M_i - \\hat{M}_c\\|^2$, where $\\hat{M}_c$ is a class-mean concept matrix updated by momentum online. This simultaneously regularizes feature invariance and logit invariance, because the matrix entries weight features by the classifier's own importance. The paper reports that this loss, with no other architectural change, gives the best average accuracy on all four benchmark settings, surpassing the strongest baseline by 1.3 to 3.7 points depending on the setting and exceeding ERM by 7 to 10 points on cross-person benchmarks.","pith_inferences":["The authors leave implicit that $L_{\\mathrm{CMS}}$ is intra-class scatter minimization in the concept space; under that reading, adding a constraint on feature or weight norms would directly test whether the gains come from genuine alignment rather than from scale shrinkage.","A testable extension is to apply the class-mean alignment per sensor axis or per body position, which could reveal which concept-matrix entries carry the domain-stable signal.","Because the momentum target is coupled to the parameters being optimized, freezing $\\hat{M}_c$ from a model pretrained on the source domains would isolate whether the target's drift matters; the paper's $\\lambda=1$ ablation is a partial step in that direction.","The same multiplicative feature-weight view appears in attention and gating layers, so concept-matrix invariance could be ported to those architectures beyond human activity recognition."],"forward_implications":["HAR models would gain several accuracy points on unseen users with only a few lines of code added to standard empirical risk minimization; the paper reports 7 to 10 point gains over ERM on cross-person benchmarks.","The regularization is model-agnostic: the authors show it also improves results when the CNN backbone is replaced with a Transformer.","Because CCIL does not need domain labels during training, it applies to mixed multi-user data where clean domain partitions are unavailable.","Consistent gains under cross-position and cross-dataset settings imply the same invariant helps when sensor placement or collection device changes, not only when the person changes."],"supporting_citations":[{"why":"Supplies the CNN backbone architecture, the DIVERSIFY baseline, and the benchmark setup the experiments follow.","marker":"Lu et al. 2024"},{"why":"The momentum update in Eq. (5) is taken from this work's momentum contrast approach.","marker":"He et al. 2020"},{"why":"Supports the premise that stable patterns persisting across domains are the causal factors worth aligning, and is cited alongside He et al. for the momentum-style update.","marker":"Chen et al. 2023"},{"why":"Provides the empirical finding that many domain-generalization methods fail to beat ERM, which motivates the new regularizer and frames the evaluation.","marker":"Gulrajani and Lopez-Paz 2021"},{"why":"Provides the DSADS dataset used for cross-person and cross-position generalization experiments.","marker":"Altun, Barshan, and Tunc ¸el 2010"},{"why":"Provides the PAMAP2 dataset used in cross-person, cross-dataset, and one-person-to-another evaluations.","marker":"Reiss and Stricker 2012"},{"why":"Provides the USC-HAD dataset used across multiple cross-domain settings.","marker":"Zhang and Sawchuk 2012"},{"why":"Provides the UCI-HAR dataset used in the cross-dataset evaluation.","marker":"Anguita et al. 2013"},{"why":"GILE is the closest prior domain-generalization method for cross-person sensor-based HAR and serves as a comparison baseline.","marker":"Qian, Pan, and Miao 2021"}],"fun_headline_variants":["Concept-matrix loss boosts activity recognition across users","New loss regularizes features and logits for HAR generalizability","Category concept invariance lifts cross-person HAR by up to 10 points","CCIL: aligning concept matrices improves sensor activity recognition","Loss that aligns class concept matrices generalizes across domains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the running per-activity mean concept matrix, updated from the model's own weights and features, is a stable and trustworthy target; if those weights or features drift or collapse, the loss can shrink without improving real generalization.","fun_headline_variants_meta":{"raw":{"variants":["Concept-matrix loss boosts activity recognition across users","New loss regularizes features and logits for HAR generalizability","Category concept invariance lifts cross-person HAR by up to 10 points","CCIL: aligning concept matrices improves sensor activity recognition","Loss that aligns class concept matrices generalizes across domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000672,"raw_usage":{"total_tokens":3071,"prompt_tokens":965,"completion_tokens":2106,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":2026}},"tokens_in":581,"tokens_out":2106,"duration_ms":15668,"temperature":1.0,"reasoning_tokens":2026,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:58:08.192442+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CCIL while tracking the rank and norm of the classifier weights $W$ and the per-class feature covariance; if $L_{\\mathrm{CMS}}$ decreases mainly through shrinking feature or weight norms, or through a low-rank collapse of $W$, the claimed alignment mechanism is not what drives the accuracy gains. A direct check is to re-run the experiments with a norm-preserving reparameterization that keeps $\\|W\\|$ and $\\|z\\|$ approximately fixed and see whether the gains survive.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the premise that stable patterns persisting across domains are the causal factors worth aligning, and is cited alongside He et al. for the momentum-style update."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the empirical finding that many domain-generalization methods fail to beat ERM, which motivates the new regularizer and frames the evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DSADS dataset used for cross-person and cross-position generalization experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PAMAP2 dataset used in cross-person, cross-dataset, and one-person-to-another evaluations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the USC-HAD dataset used across multiple cross-domain settings."},{"cited_title":"L.; et al","cited_arxiv_id":null,"evidence_quote":"Provides the UCI-HAR dataset used in the cross-dataset evaluation."},{"cited_title":"J.; and Miao, C","cited_arxiv_id":null,"evidence_quote":"GILE is the closest prior domain-generalization method for cross-person sensor-based HAR and serves as a comparison baseline."}],"review_version":1}