{"id":"95b637f7-fe61-4d20-af47-499763417e6c","arxiv_id":"2501.15070","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A time-series transformer trained with Shapley-based and contrastive pretraining outputs predictions and Shapley explanations in one forward pass, at a fraction of post-hoc explanation cost.","lead":"A new training scheme lets time-series transformers output both predictions and per-input explanations in a single pass, avoiding repeated model runs at request time. The paper reports comparable accuracy and 4-5 times faster explanations than a standard Shapley-based method on public datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (7), the stated pretraining objective, has no Shapley regression term and no dependence on the explainer head; the paper never shows how phi_gamma learns Shapley values.","rationale":"The reader's weakest assumption points at mask generalization and the ambiguous connection between Eq. (1) and Eq. (7). My concern is more upstream: Eq. (7) contains no term that supervises the explainer head at all, so the ambiguity is not merely about a schedule or sampling distribution; it is about whether the explainer ever receives a Shapley-value training signal. The KL term in Eq. (7) encourages the model to make similar predictions on original and perturbed inputs, which can be satisfied by invariant representations without performing cell-level credit assignment; it is not equivalent to minimizing the FastSHAP regression loss. This is an internal gap, not a disagreement with external consensus. I still recommend keeping the reader's CONDITIONAL verdict rather than rejecting, because the paper also defines L_gamma in Eq. (2), so a precise statement of the optimization phases, together with an ablation removing L_gamma, could resolve the issue. The reported speed advantage and prediction results are plausible in direction, but no code, error bars, or formal verification are provided, so the burden of demonstrating the Shapley-training mechanism rests with the authors.","tokens_in":8633,"tokens_out":8501,"duration_ms":81709,"concrete_test":"Run the reported pipeline exactly (run A) and an ablation (run B) in which the L_gamma terms in Eq. (2) are deleted, keeping Eq. (7), all hyperparameters, and all data splits unchanged. On 100 held-out AReM test samples, compare (i) the faithfulness scores of Table III and (ii) the Spearman rank correlation between ShapTST's output phi'_gamma and TimeSHAP's Shapley values per feature/time step. If runs A and B are statistically indistinguishable on both metrics, the Shapley regression loss is inert and the stated mechanism for Shapley-based pretraining fails. A supplementary check: symbolically differentiate L_P with respect to gamma; a zero gradient confirms the pretraining stage does not update the explainer head.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that ShapTST produces Shapley-value explanations in a single forward pass by training the explainer head with a FastSHAP-style loss. The only explicitly specified pretraining objective, Eq. (7), is L_P = InfoNCE + sum(alpha_CE * CE) + alpha_KL * KL. It contains no term of the form (v_{x,y}(S) - v_{x,y}(empty) - S^T phi_gamma(x,y))^2 from Eq. (1), and L_P does not depend on the explainer parameters gamma at all: the KL term compares predictions of the original and masked inputs, not the explainer's additive attribution to the value difference. Thus the mechanism by which phi_gamma is supervised to approximate Shapley values is absent from the stated pretraining loss. If the intended schedule is to first run Eq. (7) and then fine-tune with Eq. (2), the paper must say so and report the relative weights and phases; if L_gamma is instead optimized jointly with the backbone, the targets v_{x,y}(S) change as f_w changes, so the regression in Eq. (1) is a moving target and FastSHAP's fixed-model convergence argument does not apply. In either case, the reported faithfulness scores in Table III cannot certify that the head outputs Shapley values, because the training signal that would create them is not specified. This is more basic than the mask-generalization concern: before asking whether the head generalizes over masks, the paper must establish that the head is ever trained toward Shapley values.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ShapTST, a time-series transformer framework with an additional explainer head intended to output Shapley-value-based explanations for time steps, features, and individual cells. The stated design is to train this head with a FastSHAP-style regression loss while pretraining the backbone with a combination of InfoNCE, cross-entropy, and KL-divergence losses, and then to use the head at inference to provide explanations in a single forward pass. The authors report competitive or improved prediction performance on several classification and regression datasets, a 4-5x reduction in explanation evaluation time compared with TimeSHAP, faithfulness scores on par with TimeSHAP, and improved robustness to noise via a Shapley-based regularization term. The central proposal is that explanation and prediction can be unified during training, eliminating expensive post-hoc Shapley computation.","tokens_in":8998,"tokens_out":4541,"duration_ms":47366,"significance":"If the explainer head genuinely learns the Shapley values of the trained model, the amortized, single-pass explanation mechanism would be a valuable contribution for safety-critical time-series applications, and the idea of integrating Shapley-based supervision into pretraining is worth serious consideration. The paper usefully identifies an efficiency limitation of post-hoc Shapley computation and proposes a concrete alternative architecture. However, the manuscript currently leaves the core training mechanism for the explainer head unspecified, and several experimental details are missing, so the central claims about explanation quality and single-pass efficiency are not yet fully supported. The paper does not provide code, error bars, or hyperparameter values, which further limits verification. The potential significance is real, but the present version does not establish it.","major_comments":[{"comment":"The pretraining loss L_P in Eq. (7) has no Shapley regression term and no dependence on the explainer parameters gamma. Eq. (1) defines L_gamma as the supervised regression of the head to the marginal contributions v_{x,y}(S)-v_{x,y}(empty)-S^T phi_gamma(x,y), but Eq. (7) consists only of InfoNCE, CE, and KL terms. The text states that the KL term 'allows for Shapley value estimation,' yet that term compares predictions of the original and perturbed inputs rather than the explainer's additive attribution to the value difference. The manuscript never specifies how phi_gamma is trained toward Shapley values. If the intended procedure is a two-stage schedule (pretrain with Eq. (7), then fine-tune with Eq. (2)), the schedule and relative weights must be stated; if Eq. (1) is optimized jointly with the backbone, the targets v_{x,y}(S) change as f_w changes, so the standard FastSHAP convergence argument does not apply directly. This is load-bearing because the faithfulness results in Table III can only be interpreted as measuring Shapley accuracy if the head actually received supervision toward Shapley values.","section":"II-C, Eq. (7)"},{"comment":"The claim that the model 'simultaneously generates predictions and Shapley value estimations in a single forward pass' is not consistent with Eq. (3), which refines the head outputs by adding v_{x,y}(1) - v_{x,y}(empty) - 1^T phi_gamma(x,y)/(D*T). Computing this refinement requires evaluating v_{x,y}(empty), the model output on a fully masked input, which is not already available from the prediction pass. The paper does not explain how v_{x,y}(empty) is obtained without an additional forward pass, or whether the refinement is omitted at inference despite being part of the method description. The speed comparison in Table II therefore needs a clear statement of what exactly is measured, including whether this extra evaluation is included for ShapTST and for TimeSHAP.","section":"II-A, Eq. (3) and III-D"},{"comment":"The masking functions m_T, m_D, and m_C all rely on an 'uninformative values' matrix x~, but the paper never defines how x~ is chosen. For Shapley values, the replacement baseline is part of the definition of the value function v_{x,y}(S), so different baselines (zero, mean, per-feature median, or a learned value) lead to different explanations. Similarly, the subset distribution p(S) and the sampling weights over time, feature, and cell masks are not specified. Without these choices, the equivalence between ShapTST's head outputs and TimeSHAP's Shapley values cannot be assessed, and the faithfulness comparison in Table III is not reproducible.","section":"II-B, Eqs. (4)-(6)"},{"comment":"The abstract states that the method is validated on eight public datasets, but Table I contains six distinct datasets (AReM and Gas sensor each appear under two tasks). The paper also reports no error bars, confidence intervals, or number of random seeds for any of the prediction, faithfulness, or robustness results, and the hyperparameters alpha_CE, alpha_KL, lambda, and p_mask are introduced but never given numerical values. These omissions make it impossible to determine whether the observed prediction improvements and the robustness effects are statistically meaningful, and they prevent replication of the experiments.","section":"III, Table I and Section III-A"},{"comment":"The faithfulness score is referenced to [36] but the exact computation used to produce Table III is not described in the manuscript. It is not stated which masking procedure is used for evaluation, how the values are aggregated over time steps and features, or how the score is normalized. Since faithfulness is the main quantitative evidence for explanation quality, the reader cannot verify whether the reported numbers (e.g., 0.650 for TimeSHAP vs. 0.688 for ShapTST on AReM) are computed on an apples-to-apples basis. Please specify the faithfulness definition and the evaluation protocol.","section":"III-C, Table III"}],"minor_comments":[{"comment":"The name 'TimeShap' appears in the introduction and in the caption of Table III, while the reference is TimeSHAP [19]; please use a consistent spelling.","section":"I and III-C"},{"comment":"The dataset name 'EGG eye state' should be 'EEG eye state'.","section":"III"},{"comment":"Table II reports times without units and does not state how many samples or test instances were used for the timing comparison, nor whether the hardware and batch settings were identical for both methods.","section":"III-D, Table II"},{"comment":"In Eq. (6), S_T and S_D are vectors of different lengths, and the notation S_T * S_D is ambiguous; please clarify whether this is an outer product or a broadcasting rule.","section":"II-B, Eq. (6)"},{"comment":"The Shapley-based regularization uses a target y_target (e.g., zero) but the paper does not explain how a user would choose this target for a given feature or time step, or how the L2 loss is scaled relative to the other losses.","section":"II-D"},{"comment":"The total loss L_DT in Eq. (2) sums L_gamma over time, feature, and cell levels, but it is not explained how the three losses are weighted relative to each other and to L_w, nor how these weights are chosen in the experiments.","section":"II-A, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an early-stage preprint with a promising idea but a missing central specification: the pretraining objective in Eq. (7) does not contain the Shapley regression loss, leaving the training of the explainer head undefined. This is fixable within the scope of the paper if the authors provide the exact training schedule and loss composition, and if they strengthen the experimental reporting with error bars, hyperparameters, baseline definitions, and a precise faithfulness protocol. There is also a novelty claim about being the first to incorporate Shapley value estimation into time-series transformer training that should be checked against the wider literature before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the idea is sensible and the speed gain is believable, but the paper has a load-bearing gap—the training signal for the explainer head is never actually specified. Eq. (7), the only pretraining loss given, contains no Shapley term and no dependence on the explainer's parameters. Eq. (2) does include the FastSHAP loss, but the paper never explains how these two losses fit together. If the plan is to pretrain with Eq. (7) and then fine-tune with Eq. (2), you need to say that and give the schedule; if the explainer is trained jointly, the targets v move as f_w changes, so FastSHAP's convergence argument doesn't apply. Either way, Table III's faithfulness numbers can't certify that the head is producing Shapley values, because the training procedure that would create that behavior isn't in the paper.\n\nWhat's genuinely new is the application of amortized Shapley estimation to time-series transformers, with multi-level masking over time, features, and cells, plus a noise-robustness regularizer that uses the Shapley head. Those are legitimate contributions, and the 4-5x speedup over TimeSHAP is the kind of result that matters in practice. The prediction improvements are modest but consistent, and the robustness experiment on AReM is a nice touch.\n\nThe soft spots are mostly experimental reporting: no error bars, no seeds, the abstract says eight datasets while Table I lists six (four classification, two regression), and the faithfulness measure is internal. None are fatal, but they add up.\n\nWho is this for? Researchers working on explainable time-series models who care about inference-time cost. They'll want to see a corrected version. As it stands, the central training loop is underspecified, so I wouldn't accept it as is. But it's not a waste of anyone's time—it deserves a serious referee, with the explicit request to pin down the explainer-head training. If the authors can show that, this could be a solid paper.\n\nRecommendation: send to peer review, but expect major revision.","headline":"Plausible amortized Shapley for time-series transformers, but the training signal for the explainer head is missing from the paper.","tokens_in":9476,"tokens_out":4796,"would_cite":false,"duration_ms":40469,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ShapTST trains a time-series transformer to output Shapley-value explanations in the same forward pass as its prediction, trading post-hoc repeated inference for a one-time pretraining cost.","keywords":["time-series transformer","Shapley values","explainability","amortized explanation","pretraining","contrastive learning","masking","robustness"],"falsifier":"Compare ShapTST's per-cell Shapley estimates against exact Shapley values computed with a very high-sample post-hoc method on held-out masks that are rare in pretraining—for example masks covering large contiguous time blocks or unusual feature combinations. If faithfulness on those masks falls well below the Table III averages, the amortized estimator fits the pretraining mask distribution rather than learning the true Shapley function.","tokens_in":8455,"feed_emoji":"⚡","tokens_out":7470,"duration_ms":60493,"temperature":0.7,"pith_summary":"The paper proposes ShapTST, a training scheme that makes a time-series transformer output both its prediction and a Shapley-value explanation in one forward pass. Shapley values are per-input credit-assignment scores that say how much each time step and feature contributed to a prediction; they are normally computed after training by feeding many masked copies of the sample through the model. ShapTST instead trains an extra explainer head during a pretraining stage to imitate that repeated-inference process, so at deployment the explanation comes out with the prediction. On eight public datasets the approach matches or improves prediction accuracy, yields faithfulness scores comparable to TimeSHAP, and measures explanations 4 to 5 times faster than TimeSHAP. The practical target is safety-critical applications where every prediction must be accompanied by a reason without slowing down the system.","feed_headline":"ShapTST produces predictions and Shapley explanations in a single pass","feed_subtitle":"Training the explainer during pretraining replaces repeated inference, cutting explanation time 4-5x versus TimeSHAP.","key_machinery":"The key machinery is the explainer head paired with a Shapley-estimation loss and three masking functions. The loss supervises the head to reproduce the marginal contribution $v_{x,y}(S)-v_{x,y}(\\emptyset)$ of every masked subset, amortizing the repeated inference that post-hoc methods perform at test time; the masks $m_T$, $m_D$, $m_C$ define the subsets at time, feature, and cell granularity. A pretraining objective that sums InfoNCE, cross-entropy, and a KL-divergence between original and perturbed predictions ties this explanation learning to representation learning, and the efficiency-gap correction forces the estimated Shapley values to sum to the full prediction difference.","core_discovery":"The central claim is that Shapley explanations for a time-series transformer can be learned rather than computed post-hoc. ShapTST adds an explainer head $\\phi_\\gamma(x,y)\\in\\mathbb{R}^{D\\times T}$ trained with the FastSHAP-style loss $\\mathcal{L}_\\gamma = \\mathbb{E}_{p(x)}\\mathbb{E}_{U(y)}\\mathbb{E}_{p(S)}[(v_{x,y}(S)-v_{x,y}(\\emptyset)-\\mathbf{1}_S^\\top \\phi_\\gamma(x,y))^2]$, where $S$ ranges over subsets generated by time-, feature-, and cell-level masking functions (Eqs. (4)-(6)). After training, the model produces predictions and per-cell Shapley values simultaneously, with the estimates refined by the efficiency-gap correction of Eq. (3). The paper reports that on eight datasets this yields competitive predictions, robustness to noise, faithfulness at least on par with TimeSHAP, and a 4-5x reduction in explanation evaluation time.","pith_inferences":["The paper leaves the uninformative baseline $\\tilde{x}$ and the mask sampling distribution $p(S)$ unspecified; an editorially natural reading is that the amortized Shapley values are only guaranteed to be accurate for masks the pretraining distribution represents, so faithfulness on arbitrary held-out masks is an open empirical question.","If the estimator generalizes, the same 'pretrain an explainer head with a Shapley loss' recipe could be applied to other sequence models (speech, medical monitoring), where the savings compound because TimeSHAP-style repeated inference is per request.","The controllable Shapley target in Section II-D suggests a mechanism for injecting prior knowledge into a model—setting a known-spurious feature's Shapley contribution to zero—which could be tested as a general debiasing tool beyond noise robustness.","Since the pretraining uses only the input signal and labels, the approach could be extended to self-supervised setups with unlabeled time series, using the KL term between original and perturbed predictions as an explanation-aware augmentation."],"forward_implications":["Explanations no longer require repeated forward passes at deployment: after pretraining, ShapTST computes predictions and Shapley values simultaneously, with a measured 4-5x reduction in explanation evaluation time versus TimeSHAP on the tested datasets.","The amortized estimator makes explanation cost independent of the number of sampled subsets, so at high explanation frequencies the savings grow and may make per-request Shapley explanations practical in streaming or safety-critical settings.","The pretraining objective doubles as a representation learner: ShapTST improves over its vanilla TST baseline on several classification and regression benchmarks, with the largest AUROC gain (2.7%) on the larger Gas Sensor dataset, and improves robustness to Gaussian noise when a noisy feature's Shapley target is regularized to zero.","Because the framework is attached as an extra head plus losses, it can be layered onto existing transformer backbones; the paper demonstrates it with TST, PatchTST, and ViTST with results on par with the originals."],"supporting_citations":[{"why":"TimeSHAP, the post-hoc time-series Shapley baseline whose faithfulness and runtime are compared against.","marker":"[19]"},{"why":"FastSHAP, which supplies the Shapley-estimation loss and the efficiency-gap correction used by the explainer head.","marker":"[21]"},{"why":"SCARF, source of the InfoNCE contrastive pretraining used to learn representations from corrupted views.","marker":"[25]"},{"why":"TST, the general-purpose transformer backbone that ShapTST extends with the explainer head.","marker":"[33]"},{"why":"PatchTST, a stronger transformer baseline used to test whether the framework transfers.","marker":"[34]"},{"why":"ViTST, the vision-transformer-based time-series baseline used in experiments.","marker":"[35]"},{"why":"SpecAugment, the masking-augmentation inspiration for time/feature/cell level masks.","marker":"[24]"}],"fun_headline_variants":["Shapley explanations learned once, reused for every prediction","One-pass Shapley explanations for time-series transformers","Pretrain to explain: ShapTST cuts explanation cost 4-5x","Single forward pass gives predictions and Shapley values","Learn Shapley values during pretraining, not at inference"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the explainer head, trained on a sampled set of masked subsets, gives correct Shapley values for all masks used at test time—including masks that never appeared in pretraining—which requires the pretraining mask distribution and the uninformative replacement values to match how information is removed in deployment.","fun_headline_variants_meta":{"raw":{"variants":["Shapley explanations learned once, reused for every prediction","One-pass Shapley explanations for time-series transformers","Pretrain to explain: ShapTST cuts explanation cost 4-5x","Single forward pass gives predictions and Shapley values","Learn Shapley values during pretraining, not at inference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000275,"raw_usage":{"total_tokens":1658,"prompt_tokens":973,"completion_tokens":685,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":599}},"tokens_in":589,"tokens_out":685,"duration_ms":5243,"temperature":1.0,"reasoning_tokens":599,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:38:39.139138+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare ShapTST's per-cell Shapley estimates against exact Shapley values computed with a very high-sample post-hoc method on held-out masks that are rare in pretraining—for example masks covering large contiguous time blocks or unusual feature combinations. If faithfulness on those masks falls well below the Table III averages, the amortized estimator fits the pretraining mask distribution rather than learning the true Shapley function.","supporting_citations":[{"cited_title":"Time- shap: Explaining recurrent models through sequence perturbations,","cited_arxiv_id":null,"evidence_quote":"TimeSHAP, the post-hoc time-series Shapley baseline whose faithfulness and runtime are compared against."},{"cited_title":"Fastshap: Real-time shapley value estimation,","cited_arxiv_id":null,"evidence_quote":"FastSHAP, which supplies the Shapley-estimation loss and the efficiency-gap correction used by the explainer head."},{"cited_title":"A transformer-based framework for multivariate time series representation learning,","cited_arxiv_id":null,"evidence_quote":"TST, the general-purpose transformer backbone that ShapTST extends with the explainer head."},{"cited_title":"Time series as images: Vision transformer for irregularly sampled time series,","cited_arxiv_id":null,"evidence_quote":"ViTST, the vision-transformer-based time-series baseline used in experiments."}],"review_version":1}