{"id":"40566e88-1d77-4b52-8640-9974f3a40184","arxiv_id":"2505.09955","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"TransPL uses vector-quantized code transition matrices to generate pseudo-labels for time-series domain adaptation, outperforming prior pseudo-labeling baselines on four benchmarks.","lead":"TransPL is a new way to label unlabeled sensor data from a new user or device, by learning short code patterns from labeled data and comparing how those patterns change over time and across sensors. It reports higher accuracy than existing domain adaptation methods on four time-series benchmarks, while showing which channels drive the shift.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pseudo-label gains hinge on the untested assumption that source class-conditional code-transition matrices transfer to target classes; the paper's own Section 8 flags the channel-weighting failure mode but not this per-class transition drift.","rationale":"The reader identified the same gap: Eq. 7 applies source class-conditional transition matrices to target sequences under only a covariate-shift assumption, which does not constrain class-conditional distributions. I agree this is the most load-bearing assumption because the method's novelty and pseudo-label quality both depend on it. I also confirm the concern is empirical, not just theoretical: the per-pair tables in Appendix D show several large individual losses to SHOT that are inconsistent with the narrative of uniformly superior transition-matrix transfer. The paper's Section 8 explicitly notes the channel-weighting failure mode when informative channels are heavily shifted, but does not address the more fundamental per-class transition-drift issue. Because the paper reports a clear mechanism, strong average numbers, and a reproducible setup, I would not reject or mark it unverdictable; the fix is an additional validation step, so conditional acceptance is the right verdict and the reader's verdict stands.","tokens_in":22157,"tokens_out":1695,"duration_ms":15120,"concrete_test":"Run a controlled check: select the UCIHAR and WISDM source-target pairs where TransPL loses most to SHOT (e.g., UCIHAR 12->16, WISDM 6->19, WISDM 28->4), and compare the source class-wise transition matrices P^S_{cl,k} against the target's ground-truth class-conditional transition matrices computed from target labels (labels used only for diagnosis). Compute the per-class transition-matrix distance (e.g., average OT cost used in Eq. 6) and correlate it with per-class pseudo-label error. If target classes with large transition drift show concentrated pseudo-label errors, the transfer assumption in Eq. 7 is falsified for those regimes. If per-class drift is small on the pairs where TransPL wins and large on the pairs where it loses, the paper's averaged headline claim needs a domain-pair-level caveat.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central mechanism (Eq. 3 and Eq. 7) treats the source-trained class-conditional transition matrices P^S_{cl,k} as likelihoods for target sequences. Section 4.2 justifies transfer by citing covariate shift P(y|X)^S = P(y|X)^T. But covariate shift is a posterior invariance; it does not imply the class-conditional distributions P(X^d|y=k) are unchanged, nor that coarse-code transition probabilities within a class are domain-invariant. The likelihood in Eq. 7 is normalized only by sequence length N and is computed from source transitions, so any target-side change in a class's temporal signature systematically miscalibrates the Bayes posterior. The channel alignment weights w_d in Eq. 6 scale whole channels and cannot correct per-class transition drift within a channel. This is not merely a theoretical gap: the pseudo-label accuracy numbers in Table 3 are exactly the place where such miscalibration should show, and the reported gains are strong. The paper reports 'consistent' gains over SHOT in Table 2, but Tables 9-16 show that TransPL loses to SHOT on several individual UCIHAR and WISDM pairs (e.g., UCIHAR 12->16 Acc 66.4 vs 90.3; WISDM 6->19 Acc 48.5 vs 78.8; WISDM 28->4 Acc 54.5 vs 83.3). Those large per-pair gaps are consistent with per-class transition drift in cases where the shift is not uniform, yet the paper's central claim of averaging hides them. Since the paper's headline contribution is exactly the transferability of these transition matrices, this missing validation is the load-bearing concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TransPL, a pseudo-labeling method for time-series unsupervised domain adaptation (UDA). A VQ-VAE is trained on labeled source data with coarse and fine codebooks; from the coarse codes the method builds class-wise transition matrices per channel and channel-wise transition matrices for source and target. Target sequences are pseudo-labeled by Bayes' rule using class-conditional likelihoods from Eq. (7), a label prior, and optimal-transport-based channel alignment weights w_d from Eq. (6). The adapted model is fine-tuned with the pseudo-labeled target set. Experiments on UCIHAR, WISDM, HHAR, and PTB report mean accuracy and macro-F1 improvements over baselines, plus ablations and weak-supervision experiments.","tokens_in":22532,"tokens_out":8077,"duration_ms":73088,"significance":"If the results hold, TransPL would be a useful and interpretable new mechanism for time-series UDA, distinguished by explicit temporal-transition modeling via VQ codes and by a principled channel weighting scheme. The paper provides a reproducible code repository, includes full per-pair tables, ablations of channel alignment and weak supervision, and several focused analyses (permutation entropy of coarse vs. fine codes, channel corruption experiments) that isolate the contribution of its modules. However, the central empirical claims are weakened by the absence of variance or statistical tests, by hyperparameter selection that uses target labels, and by per-pair failures that contradict the word 'consistent' in the claimed advantage; these need to be resolved before the claimed state-of-the-art status is established.","major_comments":[{"comment":"The method's core assumption is that source class-wise transition matrices serve as target class-conditional likelihoods. The justification given in Section 4.2 follows from the covariate-shift assumption P(y|X)^S = P(y|X)^T, but that assumption only fixes the class posterior; it does not imply that P(X^d|y=k) or the coarse-code transition probabilities within class k are invariant across domains. Since Eq. (7) computes target likelihoods directly from P^S_{cl,k}, any per-class transition drift systematically mis-calibrates the Bayes posterior, and the channel weights of Eq. (6) only rescale whole channels. The paper should validate this transfer assumption, for example by comparing class-wise transition matrices estimated from labeled target test data against the source matrices and correlating per-pair distances with per-pair pseudo-label errors in Table 3, or by reporting whether the large per-pair SHOT losses in Tables 9–16 correspond to large transition drift.","section":"Section 4.2 and Eq. (7)"},{"comment":"The codebook size selection in Table 1 uses Pseudo-Label Accuracy on the target training set, which requires target labels. This is a UDA violation: the numbers of coarse and fine codes n_c and n_f, set to 8 and 64 in Section 6, are chosen with access to target ground truth, and the baseline comparisons are therefore not obtained under the same information constraint. Please re-select hyperparameters using only source validation data (as stated in the evaluation protocol) or an unsupervised criterion such as reconstruction error, and report whether the selected n_c, n_f and the other hyperparameters of Table 17 remain unchanged.","section":"Section 6 and Table 1"},{"comment":"The phrase 'consistent advantage over SHOT' is not supported at the per-pair level. For example, TransPL's accuracy is 66.4 vs SHOT 90.3 on UCIHAR 12→16, 48.5 vs 78.8 on WISDM 6→19, 54.5 vs 83.3 on WISDM 28→4, and 75.4 vs 86.5 on HHAR 4→5; the averages in Table 2 hide these large degradations. With only 10 pairs per dataset and no variance or significance test, the claim that TransPL 'demonstrates the best adaptation performance across all four datasets' is stronger than the evidence. Please report per-pair wins/losses, paired tests (e.g., Wilcoxon signed-rank), and standard deviations or confidence intervals.","section":"Section 7.1 and Tables 9–16"},{"comment":"The class-conditional likelihood computation does not specify how zero-count transitions in P^S_{cl,k} are handled. For sequences of length N≈16–20 and a codebook of 8 states, many transitions will be unobserved in the source class-wise counts, making the log-likelihood in Eq. (7) −∞ unless a smoothing or Laplace correction is applied. The paper only mentions epsilon smoothing for the channel-wise TMs in Section 5.1; please state the exact smoothing used for the class-wise TMs, since it directly affects the pseudo-label confidences and the r_top selection in Section 5.2.","section":"Eq. (7)"}],"minor_comments":[{"comment":"The '6.1% accuracy improvement, 4.9% F1 improvement' is the gain in pseudo-label accuracy from Table 3, not the final adaptation performance gain from Table 2; please state this explicitly to avoid misreading.","section":"Abstract"},{"comment":"The RAINCOAT row appears garbled ('RAINCOAT58.6 61.1 47.3 63.0 58.6 61.1 59.8 41.2'); the values seem inconsistent with the individual results in Appendix D and with the averages shown elsewhere.","section":"Table 2"},{"comment":"Since the channel weights w_d are not renormalized to sum to one, a channel with very low alignment reduces the overall pseudo-label magnitude; please specify whether this is intentional or whether the weights should be normalized.","section":"Eq. (3)"},{"comment":"The optimal transport cost matrix M in Eq. (5) is defined on coarse codes from the shared source codebook; please state explicitly that the same codebook is used for both domains and that source and target codes are aligned in index space.","section":"Section 5.1"},{"comment":"Please report how σ, τ, r_top, and the source/adaptation learning rates were selected (source validation or target labels) to allow reproducibility.","section":"Appendix E and Table 17"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the central idea is interesting. The main blockers are the target-label leakage in Table 1, the absence of per-pair statistical support for the 'consistent' advantage claim, and the unvalidated transfer of class-wise transition matrices. These are fixable with additional experiments and re-analysis, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TransPL is worth your attention if you work on time-series UDA. It introduces a new pseudo-labeling mechanism: after source training with a VQ-VAE, it builds class- and channel-wise transition matrices over coarse codes, then pseudo-labels target sequences by Bayes rule, weighting channels by optimal-transport distance between source and target channel transitions. The combination is genuinely new, and the average numbers are respectable: about 3.0% accuracy and 2.9% MF1 over SHOT across four benchmarks, with ablation evidence that both the channel alignment and the weak-supervision prior contribute.\n\nThe paper does several things well. The exposition is clear, the evaluation uses the standard AdaTime splits and a shared backbone, and the authors acknowledge the main limitation of their channel weighting in Section 8. The weak-supervision extension is a nice addition, and the pseudo-label accuracy in Table 3 is credible. I also checked the abstract's 6.1%/4.9% claim: it does match arithmetic from Table 3 when you compare against the best baseline per dataset, so the reader's flag on that point is wrong.\n\nThe soft spots are real, though. The biggest is the transfer assumption in Section 4.2. Covariate shift only constrains the posterior P(y|X); it does not imply that class-conditional code-transition matrices are invariant. The paper asserts that temporal sequences from the same class will have similar transition patterns, but supplies no direct check. This is load-bearing: the pseudo-labels are computed from source transition matrices, and channel weights cannot correct per-class drift. The stress-test note is right to call this out. The per-pair results in the appendix show TransPL losing badly to SHOT on several pairs (UCIHAR 12->16, WISDM 6->19, WISDM 28->4), which is consistent with per-class drift disrupting the likelihoods. The paper should report per-pair breakdowns of pseudo-label accuracy and ideally a per-class divergence metric between source and target transition matrices.\n\nSecond, Table 1 selects codebook size using target pseudo-label accuracy; that is a label leak in hyperparameter choice. It should be selected on source validation. Third, there are no error bars or significance tests. Ten pairs with no variance hides big swings; a paired test over the ten UCIHAR/WISDM/HHAR pairs would be easy.\n\nThis is not a fatal flaw set. The mechanism is novel and the average improvements hold up under conservatism. But the central transfer assumption needs validation before I'd trust the SOTA claim. A competent referee could reasonably ask for that as a major revision.\n\nMy recommendation: send it to peer review. It deserves referees. I would not desk-reject it. If you work on sensor UDA, it's worth a read; I would cite the method, not the headline numbers.","headline":"TransPL is a genuinely new pseudo-labeling mechanism for time-series UDA with solid average gains, but the transferability of its class-conditional transition matrices is asserted rather than shown, and the statistics need work.","tokens_in":23066,"tokens_out":4177,"would_cite":true,"duration_ms":40840,"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":"TransPL pseudo-labels time-series target data using class- and channel-wise transition matrices over vector-quantized codes, and reports the best adaptation accuracy on all four UDA benchmarks.","keywords":["unsupervised domain adaptation","time series classification","pseudo-labeling","vector quantization","transition matrix","channel alignment","optimal transport","weak supervision"],"falsifier":"Construct or select a source-target pair in which one class's temporal dynamics change (for example, an activity performed with a different rhythm or a sensor whose mounting changes) while another class keeps its pattern, and compare TransPL's per-class pseudo-label accuracy against a target-clustering baseline such as SHOT. If the shifted class's pseudo-label accuracy drops below the clustering baseline even though the channel alignment scores stay high, the central claim that source transition matrices transfer across domains is contradicted.","tokens_in":21946,"feed_emoji":"📈","tokens_out":13267,"duration_ms":111481,"temperature":0.7,"pith_summary":"TransPL is an unsupervised domain adaptation method for time series that generates pseudo-labels from explicit temporal transition statistics rather than from a black-box classifier. The paper's central claim is that modeling the source joint distribution $P(\\mathbf{X}, y)$ through class- and channel-wise transition matrices over vector-quantized (VQ) coarse codes—counting how often one learned code follows another within each class and each sensor—is enough to pseudo-label the unlabeled target domain via Bayes' rule. Each channel votes for a class through its class-conditional transition likelihood, and the votes are weighted by how well that channel's source and target transition matrices align, measured by optimal transport. On four time-series benchmarks (UCIHAR, WISDM, HHAR, PTB) the paper reports the best adaptation accuracy and macro-F1 among all compared methods, with average gains of 11.4% accuracy and 12.2% MF1 over no adaptation and 3.0% accuracy and 2.9% MF1 over SHOT. Weak supervision, in the form of a known label prior, plugs directly into the Bayes formulation and improves results further.","feed_headline":"TransPL beats SHOT on time-series UDA by counting VQ code transitions","feed_subtitle":"Class- and channel-wise Markov transitions over quantized codes give 3.0% average accuracy over SHOT.","key_machinery":"The load-bearing object is the coarse-code transition matrix: an empirical Markov transition matrix $P_{cl}^S \\in \\mathbb{R}^{K \\times D \\times n_c \\times n_c}$ built by counting transitions between vector-quantized coarse codes of time-series patches, per class and per channel. The supporting machinery is a two-level residual VQ structure (coarse codebook with $n_c=8$ codes, fine codebook with $n_f=64$ codes) that keeps transition counting tractable while preserving reconstruction; an optimal-transport channel alignment score $w_d$ that down-weights sensors whose source-to-target transition rows have drifted, using cosine distance between codes as the transport cost; and Bayes' rule with a label prior $p(k)$ that computes channel-wise class posteriors from the class-conditional transition likelihoods. Together these turn pseudo-labeling into a transparent likelihood computation: the transition matrices are the model of $P(\\mathbf{X}, y)$, the transport cost is the measure of domain shift, and the posterior is the pseudo-label.","core_discovery":"On the paper's own terms, TransPL establishes that a generative, explainable pseudo-labeling rule can outperform discriminative pseudo-labeling in time-series UDA. After training a patch encoder, decoder, coarse codebook ($n_c=8$) and fine codebook ($n_f=64$) on labeled source data, it freezes the model and infers coarse codes for both domains. From the labeled source it builds a class-wise transition matrix $P_{cl}^S \\in \\mathbb{R}^{K \\times D \\times n_c \\times n_c}$ whose entries are empirical one-step Markov transition counts between coarse codes, per class and per channel; from both domains it builds channel-wise transition matrices $P_{ch}^S$ and $P_{ch}^T$. For a target sequence it computes the log-likelihood of its coarse code sequence under each class's transition matrix (Equation 7), converts this to a channel-wise class posterior with Bayes' rule (Equation 3), and averages the posteriors across channels with weights $w_d$ derived from the earth mover's distance between source and target transition rows, with code similarity encoded in the transport cost. The paper reports this pseudo-labeling rule as the best across the four datasets, with pseudo-label accuracy gains of 6.1% (accuracy) and 4.9% (MF1) over the best baselines, and shows that the learned transition matrices are interpretable: they reveal which classes have similar temporal dynamics and which channels have shifted between domains.","pith_inferences":["An unstated consequence is that Equation (7) uses the source class-wise transition matrix as the likelihood for target sequences, so per-class drift in temporal transition patterns that is not a whole-channel shift will mis-calibrate pseudo-labels; the channel weights $w_d$ cannot repair this because they down-weight entire channels, not specific classes within a channel. A natural extension is es","The transport cost between source and target channel transition matrices is itself a direct measure of how much each sensor's dynamics changed, which could serve as a dataset-level UDA difficulty score or a sensor-selection criterion.","The same transition likelihoods could be used for anomaly or out-of-distribution detection on time series: a sequence whose code transitions are unlikely under all class transition matrices would be flagged as novel.","Because the pseudo-label is a channel-weighted posterior, the top-$r$ confidence selection in Section 5.2 could be replaced by per-sample posterior entropy thresholds, which may improve robustness for classes with sparse transition evidence."],"forward_implications":["If the reported results hold, time-series UDA pseudo-labeling no longer needs a source classifier's softmax scores or target clustering; a countable transition model over VQ codes suffices, making the pseudo-labeling process inspectable.","Weak supervision in the form of a known target label distribution can be injected as the prior $p(k)$ in Equation (3), turning a heuristic KL-divergence penalty into a principled Bayes update; the paper reports gains in all four datasets from this prior.","Because pseudo-labels are transition-based rather than amplitude-based, the method should stay accurate under amplitude-level sensor shifts as long as each class's coarse-code transition signature persists; the paper demonstrates this invariance on the ECG and HAR benchmarks.","Channel alignment weights computed from transition-matrix transport distances flag which sensors are drifting, giving a per-dataset explanation of what needs adaptation.","The coarse/fine codebook design ($n_c=8$, $n_f=64$) yields zero dead codes and tractable $8 \\times 8$ transition matrices, so the approach scales to long sequences without expensive density estimation."],"supporting_citations":[{"why":"Supplies the VQ-VAE codebook discretization that TransPL reformulates as coarse and fine codes and builds transition matrices from.","marker":"Van Den Oord et al., 2017"},{"why":"Provides the Bayes' rule and HMM-style likelihood computation used in Equations (3) and (7).","marker":"Bishop & Nasrabadi, 2006"},{"why":"Supplies the optimal-transport solver used to compute channel alignment scores.","marker":"Flamary et al., 2021"},{"why":"Defines the optimal transport and earth mover's distance formulation used in Equations (4)-(6).","marker":"Peyré et al., 2019"},{"why":"Provides the AdaTime benchmark protocol: dataset splits, source-target pairs, and evaluation used in all experiments.","marker":"Ragab et al., 2023"},{"why":"Defines weak-supervision UDA for time series and the CoDATS baseline compared in Section 7.3.","marker":"Wilson et al., 2020"},{"why":"SHOT is the strongest pseudo-labeling baseline that TransPL claims to surpass by 3.0% accuracy and 2.9% MF1 on average.","marker":"Liang et al., 2020"},{"why":"Softmax pseudo-labeling is the confidence-based method TransPL contrasts with its transition-based likelihood pseudo-labels.","marker":"Lee et al., 2013"},{"why":"Covariate shift assumption $P(y|\\mathbf{X})^S=P(y|\\mathbf{X})^T$ that justifies transferring source class-wise transition matrices to the target.","marker":"Zhang et al., 2013"},{"why":"Permutation entropy measure used to argue coarse codes capture temporal trends while fine codes capture residuals.","marker":"Bandt & Pompe, 2002"}],"fun_headline_variants":["TransPL beats SHOT on time-series UDA by 6.1% with VQ transitions","TransPL: Explainable pseudo-labels via VQ code transitions","VQ code transition matrices give explainable pseudo-labels for time-series UDA","TransPL: Markov code transitions outperform SHOT on time-series UDA","TransPL explains UDA: VQ code transitions reveal class and channel shifts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the per-class temporal transition patterns counted on the source domain remain valid for target sequences of the same class, so the source class-wise transition matrices can serve as target class-conditional likelihoods; if a class's transition signature drifts between domains while its channel marginals look similar, the pseudo-labels will be miscalibrated.","fun_headline_variants_meta":{"raw":{"variants":["TransPL beats SHOT on time-series UDA by 6.1% with VQ transitions","TransPL: Explainable pseudo-labels via VQ code transitions","VQ code transition matrices give explainable pseudo-labels for time-series UDA","TransPL: Markov code transitions outperform SHOT on time-series UDA","TransPL explains UDA: VQ code transitions reveal class and channel shifts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000989,"raw_usage":{"total_tokens":4265,"prompt_tokens":1091,"completion_tokens":3174,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":707,"completion_tokens_details":{"reasoning_tokens":3070}},"tokens_in":707,"tokens_out":3174,"duration_ms":21922,"temperature":1.0,"reasoning_tokens":3070,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:20:47.927661+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct or select a source-target pair in which one class's temporal dynamics change (for example, an activity performed with a different rhythm or a sensor whose mounting changes) while another class keeps its pattern, and compare TransPL's per-class pseudo-label accuracy against a target-clustering baseline such as SHOT. If the shifted class's pseudo-label accuracy drops below the clustering baseline even though the channel alignment scores stay high, the central claim that source transition matrices transfer across domains is contradicted.","supporting_citations":[{"cited_title":"L., Foo, C.-S., Chen, Z., Wu, M., Kwoh, C.-K., and Li, X","cited_arxiv_id":null,"evidence_quote":"Provides the AdaTime benchmark protocol: dataset splits, source-target pairs, and evaluation used in all experiments."},{"cited_title":"R., and Cook, D","cited_arxiv_id":null,"evidence_quote":"Defines weak-supervision UDA for time series and the CoDATS baseline compared in Section 7.3."},{"cited_title":"and Pompe, B","cited_arxiv_id":null,"evidence_quote":"Permutation entropy measure used to argue coarse codes capture temporal trends while fine codes capture residuals."}],"review_version":1}