acceptodds
Withdrawn from ICLR 2027

WHAT DO TIME-SERIES EXPLANATIONS ACTUALLY EXPLAIN?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

An attribution over a time-series window may indicate that a cell matters, but not in what sense. Under autocorrelation, a cell can influence the prediction almost entirely through the later cells into which its effect propagates, yet a method that ranks by association credits it in exactly the same way as it credits a cell the model reads directly. Existing faithfulness scores cannot separate these cases, because they test an explanation against a single, undifferentiated notion of importance. We characterize these effects on the Model Dependency DAG (MDDAG): we define a direct effect as a directed input–output edge on the MDDAG, and a total effect as the corresponding input–output dependency obtained when later cells are allowed to respond. Against either reference, we define OMIC, a bounded concordance score between an explainer's claimed and denied edges (effects), together with a ranking curve and its normalized area. The mediation residual serves as a diagnostic: an explainer that emphasizes mediated relevance will score highly on total effect but poorly on direct effect. We validate the references and the score on a synthetic VAR with known coefficients and apply them to known time-series explainers.

Primary area:
interpretability and explainable AI
Keywords:
Explainable AI, Time-Series Explainability, Machine Learning, Time-Series

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_PTuxOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper distinguishes direct model dependence from predictive dependence involving later time-series inputs. It defines direct and total reference effects using conditional replacements, introduces OMIC and AUOMIC to assess support and ranking consistency, and interprets the total-minus-direct residual as a mediation diagnostic. Validation includes a synthetic VAR and audits of TIMING, TimeSHAP, and KARMA on classification, regression, and forecasting tasks derived from Beijing PM2.5.

Critical strengths
  1. Relevant problem. Temporal dependence can obscure whether an attribution reflects direct model use or predictive information about other inputs. Making the intended explanation target explicit is valuable.

  2. Transparent evaluation framework. The paper states OMIC's relationship to established ranking statistics and acknowledges several limitations. Synthetic examples and public code make its claims inspectable.

Critical weaknesses

1. The effect definitions do not support the claimed mediation interpretation

The assertion that every direct edge is also a total edge (Definition 1, line 117) is false under Eqs. (2)-(6). Let S and epsilon be independent standard Gaussians, M = S + epsilon, and f(S,M) = S-M. Then Delta_dir > 0, but g_tot(v) = v-E[M|S=v] = 0, giving Delta_tot = 0. Moreover, for f(S,M) = S, the model ignores M, yet Delta_dir = sqrt(2/pi) and Delta_tot = 2/sqrt(pi), producing a positive mediation residual. This arises because the two references use different source-replacement distributions. With smaller noise, the residual ratio approaches one despite no model-mediated pathway. These exact Gaussian counterexamples undermine both the edge-inclusion claim and the interpretation of Med as a mediation diagnostic; additional assumptions or a revised interpretation are necessary.

2. The ranking metric and experimental implementation do not establish the intended evaluation claims

AUOMIC can conceal incorrect ordering within a sparse support. With 1,000 candidate edges, ten nonzero effects 1,...,10, and the correct ten-edge support ranked entirely backwards, Eqs. (10)-(12) give AUOMIC approximately 0.99398 at pi_max = 0.75, versus 1 for perfect ordering. Thus, a near-perfect score need not establish accurate within-support ranking. Separately, Eq. (6) measures distance between conditional mean responses, whereas the public delta_tot implementation averages residual-preserving paired output differences, a generally different estimand. The reported mf ratio is defined as a mean, but mediation computes the fraction exceeding 0.5. The authors should identify the revision used for Table 1, reconcile these definitions, and establish what the resulting scores actually validate.

Table 1's separate task audits and similar direct/total scores also leave the practical gain over existing faithfulness metrics unclear. The repository's baselines and planted-mediation results should substantiate that gain under the same definitions.

3. Additional submission concern regarding repository anonymity

The repository linked in Section 4.1 retains identifying contributor metadata in its public history. Specifically, commit 1aba634adb574bf927c0cec13cff3d7a092de4ad, dated September 5, 2026, 17:37:59 UTC, modifies data/raw/beijing_pm2/PRSA_data_2010.1.1-2014.12.31.csv and ts-mule/docs/make.bat. Its metadata exposes a personal-name-style commit.author.name, an email in commit.author.email, and an associated author.login distinct from the anonymous repository account. This creates a potential de-anonymization route if the contributor is a paper author.

Questions for authors
  1. What assumptions exclude the counterexamples above, and how should Med and its ratio be interpreted when effects cancel or Delta_tot = 0?

  2. Which total-effect estimand and mf statistic generated Table 1 and Figures 2-4? Please identify the corresponding code revision and reconcile it with the equations.

  3. How should AUOMIC be interpreted when the correct sparse support is internally misranked? Can targeted checks with known absent, present, and cancelling mediation establish the diagnostic claims under a consistent definition?

Reviewer_zud5OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the faithfulness of time-series explanations by distinguishing direct model dependence from total dependence that includes effects propagated through later time steps. It introduces a Model Dependency DAG framework to characterize these relationships and proposes two evaluation metrics, OMIC and AUOMIC, to assess the consistency of feature attributions with intervention-based reference effects. The framework is validated using synthetic VAR data with known dependencies and evaluated on the Beijing PM2.5 dataset across classification, regression, and forecasting tasks using TIMING, TimeSHAP, and KARMA. The results demonstrate the framework's ability to assess attribution consistency and identify differences between direct and mediated relevance.

Critical strengths

The paper highlights an important limitation in existing time-series XAI evaluation: feature importance may reflect either direct model dependence or effects propagated through temporal dependencies. Distinguishing these two forms of relevance provides a useful perspective for assessing explanation faithfulness.

The proposed framework uses conditional resampling to construct direct and total reference effects, avoiding arbitrary masking baselines. This offers a principled approach to evaluating explanations under temporal dependencies, although its reliability depends on the accuracy of the estimated conditional distributions.

The paper includes synthetic experiments with known dependencies and real-world evaluations covering classification, regression, and forecasting. This provides preliminary evidence of the framework's applicability across different time-series prediction settings, although the experimental scope remains limited.

Critical weaknesses

The paper defines direct and total effects using different conditional resampling distributions and interprets their difference as a mediation residual. However, these quantities are not established as a conventional causal effect decomposition, and the paper does not provide sufficient theoretical justification for interpreting their difference as mediated relevance. In particular, the two effects differ not only in whether subsequent variables are allowed to respond, but also in their conditional reference distributions. Consequently, a nonzero residual may reflect differences in the resampling distributions rather than genuine temporal mediation. This weakens the theoretical foundation of the paper's central distinction.

Questions for authors

The paper defines Med(e)=Δtot(e)−Δdir(e)\mathrm{Med}(e)=\Delta_{\mathrm{tot}}(e)-\Delta_{\mathrm{dir}}(e) and interprets it as a diagnostic of mediated relevance. However, the two effects use different conditional resampling distributions, and their difference may reflect changes in the reference distribution rather than mediation alone. Under what theoretical conditions can this residual reliably distinguish direct from mediated dependence?

Reviewer_45xkOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper distinguishes direct model dependence from total predictive relevance in time-series explanations using on-manifold interventions. The authors propose OMIC and AUOMIC to evaluate attribution consistency against these two references, together with a mediation residual. The framework is validated on a synthetic VAR model and evaluated using TIMING, TimeSHAP, and KARMA on the Beijing PM2.5 dataset.

Critical strengths
  • The distinction between direct and mediated relevance addresses an important ambiguity in time-series XAI, particularly under strong temporal dependencies. The unified formulation provides a useful perspective for evaluating explanation faithfulness.

  • OMIC and AUOMIC offer interpretable, bounded evaluation metrics with a meaningful chance level. The synthetic experiments provide an intuitive illustration of their ranking behavior.

Critical weaknesses
  • Definition 1 claims that every direct edge is also a total edge, which does not generally hold under the proposed definitions. For example, with independent zero-mean variables (s,m) and (f(s,m)=sm), the direct effect is positive while the mediator-averaged total effect vanishes. This raises fundamental concerns about the theoretical characterization of model dependencies and the interpretation of the mediation residual.

  • Section 3.4 claims that any subset of positive-effect edges achieves OMIC = 1, regardless of its size. However, this fails when denied edges have larger interventional effects than claimed edges. Consequently, the stated insensitivity to false negatives is not generally valid, undermining the theoretical interpretation of the metric.

  • Despite emphasizing the importance of separating direct and total relevance, the reported OMIC and AUOMIC scores are nearly identical for TIMING and TimeSHAP (Table 1). The experiments do not convincingly demonstrate scenarios where this distinction substantially changes the evaluation or ranking of explanation methods.

  • The evaluation relies on a single real-world dataset and GRU-based models, with different explainers evaluated on different tasks. Furthermore, the total-effect reference depends on conditional distribution estimation, yet its sensitivity to estimation errors is not systematically investigated. These limitations weaken the generalizability and reliability of the experimental conclusions.

Questions for authors
  • How do the authors justify the claim that every direct edge is also a total edge, given that averaging over mediators can cancel direct effects? Does the framework require additional assumptions?

  • How do the authors reconcile the false-negative insensitivity claim in Section 3.4 with cases where denied edges have larger interventional effects than claimed edges?

  • Can the authors demonstrate scenarios where direct and total evaluations lead to substantially different conclusions about explanation faithfulness, preferably by comparing multiple explainers on the same task?

  • How sensitive are OMIC and AUOMIC to conditional distribution estimation errors, and how does this affect the reliability of the mediation residual?

1 comment

Dear Authors,

Thank you for your response. I appreciate your acknowledgment of several concerns, but some fundamental issues remain unresolved.

In particular, Q1 and Q2 concern mathematically incorrect claims that require formal corrections rather than terminological clarifications. Q3 remains insufficiently addressed, as no concrete evidence demonstrates when direct and total relevance lead to different evaluations. For Q4, please clarify which supplementary experiments address estimation sensitivity and what their findings are.

I would appreciate further clarification, particularly on Q1 and Q2, before reconsidering my score.

Best,
Reviewer 45xk

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.