{"id":"a836983e-bc8d-458f-a114-ff38a7a46a19","arxiv_id":"2505.17307","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A wavelet probabilistic module, estimating densities of learned latent features and selecting among multiple smoothing scales, improves multivariate time series classification accuracy when fused with LSTM and causal CNN networks.","lead":"This paper adds a wavelet-based probabilistic feature module to an LSTM and convolutional network, and tests the combined model on 30 multivariate time series classification datasets. The new WPRCN reports the best average accuracy and rank among eight methods, with the largest gains on physiological and small-sample data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing weakness is the single-anchor density in AWPG: P(x) is a likelihood under a class-1-trained density, and the paper gives no evidence this separates non-anchor classes; the ablation attributing 6.1% to the module is confounded by fixed 5-channel pruning in A2.","rationale":"I verified the key numerical claim: summing Table 2's WPRCN column gives 2307.4/30 = 76.9%, and OS-CNN gives 2112.9/30 = 70.4%, so the reported averages are internally consistent. The table has no error bars or repeated-run statistics and no code is shipped, which already justifies the reader's CONDITIONAL verdict. The deeper technical soft spot is the one-class density assumption in AWPG, exactly as the reader identified. A density estimator trained on a single class produces likelihood values that, by construction, are most informative about that anchor class; there is no guarantee they are informative for distinguishing the remaining classes from each other. This matters because the paper's contribution and the ablation gains are attributed to these probabilistic features. I additionally note a confound in ablation A2: the APTCN channel-pruning block is fixed to C_out=5, so the 'without AWPG' condition processes raw MTS data through a severe fixed bottleneck on high-dimensional datasets (e.g. n=200 for InsectWingbeat), which can depress A2 for reasons unrelated to whether the probabilistic features are discriminative. Neither issue proves the reported 76.9% is wrong; both mean the central performance and mechanism claims are not established by the manuscript alone. The concrete test above would settle whether the single-anchor density and the pruning bottleneck are load-bearing. Since the reader already issued CONDITIONAL and my concern does not move the verdict, I keep UNCHANGED.","tokens_in":19142,"tokens_out":17520,"duration_ms":141615,"concrete_test":"Re-run the ablation on all 30 UEA datasets with two controls: (i) train AWPG on the first class, on a randomly chosen class, and on pooled classes (keeping all other components fixed), and (ii) replace A2's fixed C_out=5 pruning with a raw-data projection matched to P(x)'s 15 channels (or to the source channel count). Report average accuracy and rank for each condition. If the WPRCN-minus-A2 gap shrinks by more than about 2 points when the raw-data baseline has matched capacity, or when the anchor class is changed, then the reported 6.1% probabilistic-module gain is an artifact of the single-anchor density and the pruning bottleneck; if the gap is stable across these controls, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1.3 states that the AWPG is trained in an unsupervised way, modelling one class of the data, and Fig. 3 confirms the density is fit to the first class (y=1). The probabilistic feature sent to APTCN is P(x)_{m,j0} = \\hat{p}_I(y_E), i.e. the likelihood of the latent code under that single anchor density. For a fixed density, the likelihood can separate the anchor class from the rest, but it carries no information that distinguishes two non-anchor classes unless the anchor density happens to correlate with the true class-conditional densities in the learned latent space. The paper provides no argument or experiment establishing such a correlation; Figs. 3 and 5 only compare classes against the anchor model, and all m,j0 channels are derived from the same anchor density. The adaptive network also selects I by maximizing the anchor density, so for non-anchor inputs the chosen view is not optimizing any class-specific criterion. The main ablation evidence (A2, Table 3) does not resolve this: APTCN's first block prunes to C_out=5 channels (Section 4.2), so A2 feeds raw data through a fixed 5-channel bottleneck. The reported 5.8% drop (and the 6.1% drop for A1) could therefore reflect the arbitrary anchor choice plus the information loss from pruning the raw input, rather than the discriminative value of the probabilistic features. If the anchor density fails to separate non-anchor classes, the central claim that the wavelet probabilistic module drives the improvement is unsupported, even if the average-accuracy numbers in Table 2 are arithmetically correct.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Wavelet Probabilistic Recurrent Convolutional Network (WPRCN) for multivariate time series classification. The architecture runs in parallel an LSTM module, a causal fully convolutional network with squeeze-and-excitation, and a novel probabilistic module composed of an Adaptive Wavelet Probabilistic Feature Generator (AWPG) and a Channel Attention-based Probabilistic Temporal Convolutional Network (APTCN). AWPG maps the input to a latent space with a GRU encoder-decoder, estimates a wavelet density under multiple forgetting factors, and uses an adaptive network to select one view; APTCN prunes the channels, applies efficient channel attention, and processes the probabilistic features with a dilated causal TCN. The combined features are fused and classified by a softmax layer. The method is evaluated on 30 UEA datasets against seven benchmarks, reporting the best average accuracy (76.9%) and average rank (2.5). Ablations that remove the probabilistic module, the feature generator, and the channel attention report drops of 6.1, 5.8, and 4.3 percentage points, respectively.","tokens_in":19540,"tokens_out":8446,"duration_ms":91521,"significance":"If the central mechanism were validated, WPRCN would be a useful extension of wavelet density estimators to deep multivariate time series classification, and the proposed probabilistic module could be plugged into other architectures. The paper has clear strengths: evaluation on a large external benchmark (30 UEA datasets) against seven competitors, an ablation study, and a well-described modular design. These strengths make the empirical phenomenon worth investigating. However, the paper does not currently establish that the probabilistic features themselves carry class-discriminative information beyond the anchor class, and the ablation evidence is partly confounded by the fixed channel-pruning bottleneck. The statistical significance of the headline comparison is also not demonstrated. As a consequence, the significance would be moderate rather than transformative if the mechanism were verified; at present the claims outrun the evidence.","major_comments":[{"comment":"The paper explicitly states that the AWPG is trained in an unsupervised way, modelling one class of the data, and Eq. (5) optimises a one-class threshold beta. The probabilistic feature P(x)_{m,j0} is therefore the likelihood of the latent code under a density estimated from a single anchor class. For a fixed density, this likelihood can separate the anchor class from the rest, but it carries no information that distinguishes two non-anchor classes unless the anchor density is correlated with the true class-conditional densities in the learned latent space. Figures 3 and 5 only compare all classes against the anchor model; they do not show whether non-anchor classes are separated from each other. The adaptive network also selects I by maximising the anchor density, so for non-anchor inputs the selected view is not optimised for any class-specific criterion. The central claim that the wavelet probabilistic module drives the improvement therefore needs direct evidence. Concretely, I would like to see a measure of class separability of P(x) among non-anchor classes, an ablation in which the density is estimated on all classes or per class, and a comparison with a non-informative probabilistic feature of the same channel count.","section":"Section 4.1.3"},{"comment":"Ablation A2 removes only the AWPG but keeps the APTCN, which always reduces its input to Cout=5 channels in the channel pruning block (Section 4.2). When the APTCN receives raw source data rather than P(x), the fixed 5-channel pruning can discard discriminative information, especially for datasets with many input channels (for example, InsectWingbeat has n=200). The reported 5.8% drop between A2 and WPRCN is therefore confounded: it may reflect information lost by pruning the raw input rather than the value of the probabilistic features. A proper control should keep the APTCN architecture identical and compare P(x) against raw data with the same number of channels, or against probabilistic features produced by a trivial density estimator.","section":"Section 4.2 / Table 3"},{"comment":"The paper reports a single accuracy per dataset with no error bars, number of runs, or seeds. The CD diagram in Fig. 6 is presented without the critical difference value or significance bars; given 30 datasets and 8 algorithms, the standard Nemenyi CD is about 1.9, while the rank differences between WPRCN and OS-CNN (1.1) and between WPRCN and MLSTM-FCN (1.4) are smaller than that. Thus the claim that WPRCN statistically outperforms all benchmarks is not supported by the evidence as reported. Please report standard deviations over runs, the number of runs, and pairwise significance tests, or soften the claim accordingly.","section":"Section 5.1 / Fig. 6"},{"comment":"The paper repeatedly claims robustness to non-stationarity and noise, but the experimental protocol does not directly test either. The pool of forgetting factors alpha_Gamma={1,1/10,1/100,1/500,1/1000} is fixed and no justification is given that these five values cover the relevant rates of change across all datasets. The UEA benchmark is a static archive; no dataset with controlled non-stationarity or injected noise is used. Please add targeted experiments, or explicitly restrict the non-stationarity and noise claims to the physiological datasets and to the qualitative behaviour of Eq. (2).","section":"Section 3.3 / Section 5"}],"minor_comments":[{"comment":"The first sentence of Section 5.1 contains the typo WRPCN; it should be WPRCN.","section":"Section 5.1"},{"comment":"The sentence ending with classification performance,e contains an extraneous e and should be corrected.","section":"Section 5.2"},{"comment":"In the manuscript text, Fig. 6 appears only as a placeholder sequence of rank labels and not as a rendered CD diagram; please include the actual figure with the CD value and the significance bars.","section":"Fig. 6"},{"comment":"The relationship between the channel pruning output Cout=5 and the later statement that the APTCN channel output is chosen between 20 and 25 is confusing; please clarify which hyperparameter each value refers to.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is built directly on the authors' own wavelet density estimator from references [19] and [40], and the evaluation does not include an independent reproduction of that estimator. This is not disqualifying, but the editor may wish to ensure that a reviewer with expertise in wavelet density estimation checks whether the closed-form estimator is applied within its validity. In addition, no code or run configuration is released, which makes it difficult to verify the absence of information leakage in the adaptive model selection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth engaging with, but the load-bearing claim needs more evidence than the paper currently provides. What is genuinely new is the wavelet probabilistic module (AWPG + APTCN) explicitly built for MTSC, with an analytic WDE that avoids the usual computational bottleneck, wrapped in GRU encoding and channel attention. The 30-dataset UEA evaluation is broad, the average rank of 2.5 against seven baselines is solid, and the ablation shows a consistent internal story: removing the whole module costs 6.1%, removing just the generator costs 5.8%, removing attention costs 4.3%. The paper also does something rare: it openly discusses where it loses to DA-Net, OS-CNN, CMFM-RF, and MLSTM-FCN in Section 5.2.\n\nThe soft spot is the one the stress-test note puts its finger on. Section 4.1.3 says the AWPG is trained unsupervised, modelling one class of the data, and the probabilistic feature sent to APTCN is the likelihood under that single anchor density. Nothing in the paper demonstrates that this likelihood separates non-anchor classes from each other; Figures 3 and 5 only compare classes against the anchor model. The adaptive network also picks the view by maximizing the anchor density, so for non-anchor inputs the selection is not optimizing any class-specific criterion. On top of that, the ablation A2 is confounded: APTCN's first block prunes to five channels, so feeding raw data through that bottleneck loses information regardless of whether the probabilistic features carry signal. The reported 5.8% drop could be an artifact of the arbitrary anchor plus pruning, not evidence for the probabilistic module.\n\nOther issues are more routine: no error bars or number of runs, no code, no direct non-stationarity experiment, and the CD diagram does not show pairwise significance against the closest competitors. The self-cited WDE is not a flaw by itself, but independent reproduction of that component would help.\n\nWho is this for? Someone working on MTSC architectures, especially with physiological data, who wants a new modular component to test. The paper is not a breakthrough, but it is a serious engineering effort with a plausible empirical win. Right now the central mechanism is under-supported, but every problem is addressable in revision: run the ablation with a per-class density or vary the anchor, report runs and significance, and release code. I would send it to reviewers, with a request for those experiments rather than a desk rejection.","headline":"Plausible, extensively evaluated architecture, but the central probabilistic module rests on a single-anchor density whose discriminative value for non-anchor classes is never shown.","tokens_in":20033,"tokens_out":1908,"would_cite":false,"duration_ms":15788,"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 adding an adaptive wavelet density-estimator module to an LSTM and a causal CNN raises multivariate time series classification to 76.9% average accuracy, ahead of all seven benchmark methods.","keywords":["multivariate time series classification","wavelet density estimation","probabilistic neural networks","temporal convolutional networks","channel attention","non-stationary time series","LSTM","time series benchmark"],"falsifier":"Run the same 30-dataset benchmark with the AWPG trained on a different class, or on a mixture of all classes, instead of the first class; if average accuracy drops to the level of the module-free ablation, the improvement is an artifact of the anchor-class choice rather than a general property of the probability-domain features.","tokens_in":18937,"feed_emoji":"📈","tokens_out":8610,"duration_ms":63733,"temperature":0.7,"pith_summary":"The paper sets out to show that probabilistic features taken from the density of a time series, rather than from its raw waveform alone, improve multivariate time series classification in exactly the regimes where deep models struggle: scarce training data, noise, and non-stationarity. It instantiates this idea as WPRCN, a network with three parallel branches: an LSTM, a causal fully convolutional network, and a wavelet probabilistic module that generates density features and analyzes them with a temporal convolutional network. On the 30-dataset public benchmark the model reports 76.9% average accuracy and an average rank of 2.5, against 70.4% for the best comparator, and the authors' ablation tests attribute a 6.1-point accuracy drop to removal of the probabilistic module. The underlying justification is that a wavelet density estimator with a closed-form, constant-complexity update makes the probability domain computationally affordable inside a deep network.","feed_headline":"Wavelet probability module beats 7 classifiers at 76.9%","feed_subtitle":"An adaptive wavelet density estimator fused with LSTM and causal CNN outranks every baseline on 30 datasets.","key_machinery":"The load-bearing component is the multi-receptive-field wavelet probabilistic network (MRWPN), built on a radial B-spline wavelet density estimator whose coefficients update in closed form as $\\hat{w} = (1-\\alpha_\\Gamma)\\,\\hat{w} + \\alpha_\\Gamma\\, 2^{\\frac{nj_0}{2}} \\varphi\\left(\\|2^{j_0} x - k\\| + \\frac{m}{2}\\right)$, where $\\alpha_\\Gamma$ is a set of forgetting factors that define different temporal receptive fields and $m$ and $j_0$ control smoothness and resolution. This one-pass update is what makes a density model cheap enough to sit inside a deep network. The adaptive wavelet probabilistic feature generator (AWPG) wraps this estimator in a GRU encoder-decoder to form a latent space, produces $|\\Gamma|$ density views of one anchor class, and uses a softmax adaptive network to pick the view with the highest density; the resulting feature $P(x)_{m,j_0}$ aggregates every combination of $m$ and $j_0$. The channel-attention probabilistic temporal convolutional network (APTCN) then prunes the channels, enhances them with efficient channel attention, and models them with a dilated causal TCN.","core_discovery":"The central claim is that the probability-domain representation generated by a wavelet density estimator carries class-discriminative structure that ordinary temporal features miss, and that this representation can be computed cheaply enough to fuse with deep sequence models. WPRCN realizes the claim with three parallel branches: the LSTM and C-FCN extract temporal features, while the probabilistic branch maps the multivariate input into a low-dimensional GRU latent space, fits an ensemble of wavelet density estimates with different forgetting factors and smoothness settings to that space, and uses a small adaptive network to select the most informative density view for each input. The selected features are then pruned, re-weighted by channel attention, and analyzed by a dilated causal temporal convolutional network, and all three branch outputs are concatenated for classification. On the 30 benchmark datasets, WPRCN reports 76.9% average accuracy and 2.5 average rank, beating every benchmark while winning or tying on 14 of the datasets and posting its largest margins on the smallest training sets. Ablation results are read as showing that the probabilistic module, not the LSTM and CNN backbones alone, produces most of the gain.","pith_inferences":["A cheap stress test the paper does not run is to vary the anchor class; if the 6.1-point ablation gain depends on which class is modelled, the method's robustness to class imbalance and label ordering is weaker than the headline accuracy suggests.","The same ensemble of forgetting factors is naturally an online estimator, so the module could be applied to anomaly detection or forecasting on non-stationary streams, uses the paper mentions only briefly.","The visualizations suggest the selected density view could serve as an interpretable summary of which temporal scale dominates each input, although the paper does not evaluate interpretability quantitatively."],"forward_implications":["If the reported results hold, the wavelet probabilistic module can be attached to an existing multivariate time series classification model as a parallel branch and should improve accuracy without changing the host model's own feature extractors.","Because the density estimator updates in one pass without backpropagation, the probabilistic features are inexpensive to add, pointing toward streaming or low-compute deployments.","The module's smoothness and resolution parameters adapt to training-set size, so the method is specifically useful for small-sample classification problems.","The architecture-agnostic design implies the same probabilistic features could be combined with backbones other than LSTM and C-FCN, such as transformers or graph networks."],"supporting_citations":[{"why":"Supplies the closed-form, constant-complexity wavelet density estimator that the MRWPN is built on.","marker":"[40]"},{"why":"Establishes the generative wavelet-based probabilistic neural network and the analytic WDE that WPRCN extends to deep architectures.","marker":"[19]"},{"why":"Provides the LSTM/FCN fusion baseline and the configuration used for WPRCN's LSTM and C-FCN branches.","marker":"[12]"},{"why":"Is the strongest benchmark in the comparison, at 70.4% average accuracy, that WPRCN claims to outperform.","marker":"[27]"},{"why":"Supplies the 30 multivariate time series datasets and the evaluation protocol used in the experiments.","marker":"[43]"},{"why":"Provides the dilated causal TCN architecture and key hyperparameters used by APTCN.","marker":"[32]"}],"fun_headline_variants":["Wavelet probability module beats all benchmarks on 30 sets","Probabilistic wavelet features win on scarce time-series data","Adaptive wavelet estimator outperforms 7 classifiers","Wavelet density fusion lifts time-series accuracy to 76.9%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the single class chosen to train the density estimator has a distribution that differs from the other classes in every dataset; if that anchor-class density does not systematically separate classes, the fused probabilistic features add no discriminative signal and the reported gain would disappear.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet probability module beats all benchmarks on 30 sets","Probabilistic wavelet features win on scarce time-series data","Adaptive wavelet estimator outperforms 7 classifiers","Wavelet density fusion lifts time-series accuracy to 76.9%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000659,"raw_usage":{"total_tokens":3053,"prompt_tokens":1026,"completion_tokens":2027,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":1971}},"tokens_in":642,"tokens_out":2027,"duration_ms":12371,"temperature":1.0,"reasoning_tokens":1971,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:48:37.570221+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 30-dataset benchmark with the AWPG trained on a different class, or on a mixture of all classes, instead of the first class; if average accuracy drops to the level of the module-free ablation, the improvement is an artifact of the anchor-class choice rather than a general property of the probability-domain features.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the closed-form, constant-complexity wavelet density estimator that the MRWPN is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the strongest benchmark in the comparison, at 70.4% average accuracy, that WPRCN claims to outperform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the dilated causal TCN architecture and key hyperparameters used by APTCN."}],"review_version":1}