TIMING: Rewriting the Path for Time-Series Explanations

0. Source and Link Table

The links below were accessed on 2026-08-30. The arXiv page shows v1, 2025-06-05; no later version was found. The paper has no separate data repository. MIMIC-III requires PhysioNet credentials and a DUA; the public repository provides processing scripts for the other datasets and checkpoints except for MIMIC-III.

1. Motivation: The Problem Is Not Unsmooth Attribution, but Evaluation That Erases Direction

A time-series explanation typically assigns an importance score to every (t,d) location. With signed attribution, a positive value means that the point raises the predicted-class probability, while a negative value lowers it. Common evaluations mask the top-K points at once and then measure changes in accuracy or target-class probability. The paper identifies two confounds.

First, explaining the model and explaining the label are conflated. If a model has learned a spurious correlation, masking an erroneous feature that the model actually uses may improve classification accuracy; treating a larger accuracy drop as greater faithfulness then mistakes agreement with ground truth for faithfulness to model behavior. Second, directional cancellation. If the top-K set contains points that both raise and lower the prediction, removing them together can yield almost no net change even when the ranking is correct. Conversely, a worse-ranked set whose effects share one direction can produce a larger net change.

The MIMIC-III preliminary experiment quantifies this directly: in Table 1, IG-Unsigned has CPD(K=50) of 0.342±0.021, whereas IG-Signed has 0.248±0.010. Yet after masking 10%, the accuracy of retaining the original prediction (lower is better) is 0.855±0.011, substantially better than IG-Unsigned at 0.974±0.001. The metrics are not merely slightly different; they reverse the method ranking (paper p. 3, Table 1).

The authors further note that the original IG path αx scales all time steps simultaneously. It preserves relative temporal relationships throughout, but may also traverse regions outside the training distribution. Thus, if the crucial evidence is how the model reacts when a contiguous segment is disrupted relative to its neighborhood, a single straight-line IG path cannot reveal that conditional change (pp. 4–5, §4.2).

2. Narrative, Assumptions, and Trade-offs

Rather than introducing a new explainer immediately, the paper develops its argument in three steps:

1. show that existing evaluations can misjudge signed attribution, and introduce CPD and CPP;
2. show that plain IG already outperforms many time-series-specific methods under the new metrics;
3. then randomize the IG baseline/path and preserve original values in contiguous segments rather than independent points, yielding TIMING.

The central assumption is that a contiguous segment is an appropriate minimal unit of temporal semantics. This respects local dependence more than independent Bernoulli masks, but it is not learned from data; s_min,s_max,n are manually specified, and each segment selects only one channel. The method therefore favors single-channel contiguous events, with no structural guarantee for synchronized cross-channel events, discontinuous periodic patterns, or long-range causal relationships.

The trade-off is explicit: TIMING gains multi-context attribution but loses IG completeness. Proposition 4.4 acknowledges that the sum of all attributions is no longer guaranteed to equal F(x)-F(x') (p. 5). In other words, TIMING values no longer form a strictly conserved allocation of the prediction difference; they are conditional averages of local gradients across randomized contexts.

3. Equation-Level Formalization

Let the classifier be F: R^(T×D) → [0,1]^C, with predicted class ŷ=argmax_c F_c(x). Standard IG for point (t,d):

IG_t,d(x) = (x_t,d-x'_t,d) ∫_0^1 ∂F_ŷ(x'+α(x-x'))/∂x_t,d dα.

The paper primarily uses a zero baseline. Let binary mask M_t,d=1 indicate that the point moves from 0 to the input along the path, while M_t,d=0 means the original value is retained throughout the path:

z(α;M) = α(M⊙x) + (1-M)⊙x
MaskingIG_t,d = x_t,d M_t,d ∫_0^1 ∂F_ŷ(z(α;M))/∂x_t,d dα.

RandIG takes a conditional expectation over independent Bernoulli masks; TIMING instead uses the contiguous-segment generator G(n,s_min,s_max) to define the mask distribution:

TIMING_t,d(x) = E_{M~G}[MaskingIG_t,d(x,M) | M_t,d=1].

The computational device in Proposition 4.1 replaces multiple complete IG paths with one random mask sampled at each α, approximating the inner expectation with a single Riemann sum. It preserves Sensitivity and Implementation Invariance (Propositions 4.2–4.3), but not completeness (Proposition 4.4). The paper’s phrase “maintaining its theoretical properties” therefore needs qualification: two key axioms are preserved, not the entire IG axiom set.

The new metrics mask points in descending order of |A| and denote x↑k as the input after removing the first k high-attribution points:

CPD(x)=Σ_{k=0}^{K-1} ||F(x↑k)-F(x↑(k+1))||_1 (higher is better)
CPP(x)=Σ_{k=0}^{K-1} ||F(x↓k)-F(x↓(k+1))||_1 (lower is better)

CPD measures the cumulative movement of the output as important points are removed one by one; CPP applies the same operation to the lowest-attribution points. Taking an L1 norm at every step prevents positive and negative effects from directly cancelling in a single top-K mask, but it also accumulates path length rather than endpoint displacement, so oscillations in model output are counted repeatedly.

Input x, frozen classifier F, number of segments n, length range [s_min,s_max], integration steps n_samples
ŷ ← argmax F(x); TotalGrad ← 0; count ← 0
for α = 0, 1/n_samples, ... < 1:
    Construct an interpolation point from the zero baseline
    Randomly choose n contiguous (channel, start, length) segments
    Restore those segments to their original values in x and record which positions were not restored
    Compute the input gradient of F_ŷ; restored-segment gradients do not enter that point’s conditional average
    Accumulate gradients and valid counts
Return (accumulated gradient ⊙ x) / valid count per point

4. Key Mechanisms and Engineering Details

Code inspection also shows that the six real-world datasets use different segment settings: Freezer uses (5,10,100), Wafer (5,10,152), Boiler (50,1,36), PAM (10,10,600), Epilepsy (10,10,10), MIMIC (50,10,48) (official repository scripts/real/run_10perc_masking_our.sh). Thus, the claim that the same method needs no dataset-specific tuning does not hold; Table 6 conducts sensitivity analysis only on MIMIC.

5. Datasets, Tasks, and Metrics

Table 7 (Appendix p. 15) lists eight benchmarks: Switch-Feature 1,000×100×3, State 1,000×200×3; MIMIC-III 22,988×48×32, PAM 5,333×600×17, Epilepsy 11,500×178×1, Boiler 90,115×36×20, Wafer 7,164×152×1, and Freezer 3,000×301×1. Tasks cover synthetic saliency, ICU in-hospital mortality, activity recognition, EEG seizure detection, boiler faults, wafer anomalies, and freezer-source classification.

For synthetic data, AUP/AUR can be compared against the generator-defined saliency map, but this assumes that the trained model actually uses the points specified by the generator. Real data have no ground-truth explanation, so CPD/CPP, Acc, CE, Sufficiency, and Comprehensiveness are all proxies. A high CPD only means that replacing points one by one in the given order causes large cumulative output changes; it does not mean clinicians regard the explanation as correct or that highlighted variables are causal risk factors.

6. Complete Main-Result Matrix

7. Layered Ablations, Hyperparameters, Phase Changes, and Statistics

Table 5 (p. 8) decomposes IG→RandIG→TIMING: MIMIC Zero CPD moves from IG at 0.342±0.021, to RandIG(p=.3/.5/.7) at 0.350/.353/.354±approximately .022, and then TIMING at 0.366±0.021. The randomized baseline contributes about +0.012, while contiguous segments add another +0.012 over the strongest RandIG; the two gains are comparable, so the full improvement cannot be attributed to temporality. Under Avg replacement, TIMING only moves from IG at 0.172 to 0.177。

Table 6 (p. 8) gives 12 MIMIC hyperparameter settings. Zero CPD ranges from 0.345–0.366, an absolute span of 0.021; Avg ranges 0.173–0.177, whose actual span is 0.004. The text states that the maximum difference in average performance is 0.04, inconsistent with the table; a misplaced decimal point is a reasonable inference. Larger s_max=48 generally outperforms s_max=10, but there is no abrupt phase transition; n from 10 to 100 is also non-monotonic.

The main table reports mean±standard error over five-fold cross-validation, which is a strength; however, the paper does not report paired significance tests, confidence intervals, effect sizes, or multiple-comparison corrections. Freezer’s +0.004 gain, with SE around 0.11 for each method, therefore cannot establish statistical significance from the table. Even though the Boiler gap far exceeds the SE, visual separation is not a formal test. If methods share the same folds, a paired test by fold or a paired bootstrap difference would be appropriate, with multiple-comparison control across the 12 columns.

8. Computational Efficiency

Figure 4 (p. 9) reports seconds per sample and CPD over the full MIMIC test set: DeepLIFT 0.017 s, 0.142; IG 0.034 s, 0.342; GradSHAP 0.034 s, 0.327; TIMING 0.040 s, 0.366; FO 2.948 s, 0.016; AFO 3.540 s, 0.120; LIME 3.742 s, 0.071. TIMING is about 17.65% slower than IG, with CPD about 7.02% higher; it is about 88.5× faster than AFO. Because the authors exclude methods that require post-training from this figure, it does not support the conclusion that TIMING is faster than all time-series XAI methods.

9. Core Contributions, Reusable Design, and Transfer Potential

There are two substantive contributions. The first may be more important than TIMING itself: CPD/CPP replace masking K points at once with a stepwise path, directly exposing cancellation in signed attribution. This can transfer to graph nodes, image patches, and text tokens wherever a replacement operation can be defined. The second contribution places a structured baseline distribution inside the IG path: instead of training another explainer, domain structure is encoded through the input path. The idea can extend to spatially connected regions, frequency bands, event windows, or synchronized multivariate blocks.

Transfer should preserve three design disciplines: structural units must be supported by domain assumptions or data; replacement should approximate the conditional data distribution; and validation should not rely only on metrics introduced with the method, but should include external ground truth, counterfactual consistency, human tasks, or downstream debugging benefit.

10. Related Work: Problem Origins and Technical Lineage

┌─ Gradient-axiom line ─ IG (2017) ───────────┐
Model-agnostic local XAI ─┤                                          ├─ TIMING (2025)
                         └─ Temporal perturbation ─ Dynamask (2021) ─┤   ├─ CPD/CPP: stepwise evaluation
                                               ├─ ContraLSP (2024) ─┤   └─ Random contiguous-segment paths
                                               └─ TimeX++ (2024) ───┘
                                                         │
                                  OOD / signaling / cancellation as turning points

11. Limitations Stated by the Paper

1. Evaluation can be manipulated through signed ordering. Section 3.2 (p. 4) explicitly states that a signed method could raise CPD/CPP by alternating positive and negative points in absolute-value order, so signs should not be used to manipulate ranking.
2. Completeness is lost. Proposition 4.4 and the subsequent discussion (pp. 5–6) explicitly state that the attribution sum no longer equals the output difference.
3. The reproduction environment differs. The official README states that the public training scripts may not fully reproduce the paper because they use a different training environment; the MIMIC checkpoint is also unavailable due to access restrictions.

The paper has no standalone Limitations section and reports no user study, causal validation, or deployment test.

12. Independent Critique (Evidence → Inference)

1. [Metric construct validity] Because CPD sums stepwise L1 output changes and obeys the triangle inequality CPD ≥ ||F(x)-F(x↑K)||₁, oscillations are counted repeatedly. High CPD therefore need not identify net contribution more accurately; the masking trajectory may simply be more circuitous.
2. [Circular validation] Because the core method and its primary winning metric are introduced in the same paper, while Extrmask is clearly stronger on existing Acc/CE/Comp measures (Table 2: Comp×100 23.370 versus 0.436), a clean sweep on CPD alone cannot establish generally superior explanation quality. External tasks or independent ground truth are needed.
3. [Statistics] Freezer Zero moves only from IG .405±.111 to .409±.109, a 0.004 difference far below either method’s SE. The point-estimate fact of improvement on all six datasets therefore cannot establish a reliable Freezer improvement.
4. [Attribution conservation] Because TIMING explicitly lacks completeness, total attribution is not fixed across samples or segment hyperparameters. Cross-sample comparisons of absolute attribution magnitude therefore lack IG’s original scale.
5. [Missing OOD evidence] The paper argues that partial retention mitigates OOD behavior, but reports no path-sample density, discriminator AUC, nearest-neighbor distance, or calibration. Being more in-distribution is therefore a plausible mechanism hypothesis, not a directly validated empirical conclusion.
6. [Structural bias] Because each random segment selects only one channel, PAM’s synchronized 17-channel actions and Boiler’s 20-sensor coupling are not explicitly grouped. The method may fragment cross-channel events, and its large Boiler gain does not show that it learned physical coupling.
7. [Hyperparameter extrapolation] Because scripts for the six datasets use different (n,s_min,s_max), while sensitivity is reported only on MIMIC at T=48, comparable robustness cannot be inferred for PAM at T=600 or Freezer at T=301.
8. [Reproducibility] Because the README explicitly warns that the training scripts do not fully reproduce the results and the MIMIC checkpoint is unavailable, third parties without data access cannot verify the central medical-results table end to end. This limits auditability.

13. Open Challenges: Actionable Head-to-Head Experiments

14. Recent-Work Radar

1. Optimal Information Retention / ORTE, ICML 2025: uses conditional mutual information to unify redundancy and completeness, directly addressing TIMING’s lack of a global optimization criterion; it still depends on mask learning and mutual-information approximation.
2. Start Smart / StartGrad, ICLR 2025: initializes mask optimization with gradients, connecting fast gradient methods with high-quality perturbation methods; it is a strong equal-budget baseline.
3. DeltaSHAP, 2025-07-03 arXiv: explains changes in online patient risk rather than a static score, making directionality and real-time behavior more relevant to clinical monitoring.
4. Cross-domain Integrated Gradients, 2025-05-19 arXiv: extends IG to frequency, ICA, and decomposition domains while preserving completeness; it shows that TIMING’s contiguous temporal segment is not the only reasonable semantic unit.
5. TimeSAE, 2026-01-14 arXiv: uses sparse decoders to find interpretable temporal concepts, moving point attribution toward concept-level explanations; faithfulness still requires validation against black-box behavior.
6. Delta-XAI, OpenReview: attempts to unify explanations of prediction evolution. It continues TIMING’s focus on directional information but must avoid creating another self-favoring evaluation loop.

15. Conclusion

The most valuable part of TIMING is not another time-series attribution method, but its dissection of the evaluation protocol: masking the top-K points together can indeed cancel positive and negative evidence. Methodologically, randomized contiguous-segment paths provide an inexpensive, model-agnostic, transferable structural prior. TIMING leads 12/12 CPD cells across six real-world datasets, with Boiler Zero improving over IG by +109.84%, while increasing MIMIC inference only from IG’s 0.034 s/sample to 0.040 s/sample。

The strict conclusion must nevertheless be narrower: TIMING does not win on every existing metric, does not guarantee completeness, does not directly show that its paths are more in-distribution, and provides no user-based or causal validation. The most defensible characterization is:it is a strong baseline and research starting point for fair evaluation of signed attribution plus structured IG paths, not a method that completes the loop of clinically trustworthy explanation.