Pith. sign in

REVIEW 3 major objections 6 minor 7 references

L2GTX: From Local to Global Time Series Explanations

T0 review · 3 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Local cues merge into faithful global time-series explanations

desk verdict A sensible, carefully described pipeline for aggregating local event-based explanations into global time-series summaries, but the headline metric measures local surrogate fidelity, not the quality of the aggregated global explanation, so the central claim is unsupported as it stands. read the letter →

arxiv 2603.13065 v2 pith:WGOAESF3 submitted 2026-03-13 cs.LG cs.AI

Ephrem Tibebe Mekonnen , Luca Longo , Lucas Rizzo , Pierpaolo Dondio This is my paper · ORCID
classification cs.LGcs.AI
keywords ExplainableAItimeseriesclassificationglobalexplanationslocal-to-globalaggregationeventprimitivesmodel-agnosticinterpretabilitysurrogatefidelityclusterconsolidation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that faithful global explanations of a black-box time-series classifier can be built by aggregating local explanations from a small, representative set of instances. It proposes L2GTX, which extracts clusters of parameterised temporal event primitives (trends and extrema) from instance-level explanations, merges similar clusters across instances, selects the most cluster-covering instances under a budget, and summarises the events per cluster. The claim is that this yields concise, interpretable, class-wise global explanations, and that consolidation does not degrade faithfulness, measured as mean local surrogate fidelity (R²). A sympathetic reader would care because global explanations for time series are rare, mostly model-specific, and usually lack temporal structure; L2GTX aims to fill that gap without accessing model internals.

What carries the argument

The load-bearing machinery is the instance–cluster importance matrix M together with a coverage-weighted greedy selection. Each local explanation assigns importance weights to clusters of parameterised event primitives (trends and extrema); merging maps instance-specific clusters into global clusters; M records the importance of each global cluster for each instance; a SP-LIME-style global importance score I_j = sqrt(Σ_i |M_ij|) ranks clusters; and a greedy algorithm chooses, within a budget B, the instances that add the most globally important yet uncovered clusters. The aggregated events from the covered clusters are then summarised by mean ± std of their start time, duration, time, and va

What would settle it

Compute Global Faithfulness after replacing the geometric cluster merging with random assignment of local clusters to global groups (or random the instance selection) while keeping the same number of global clusters. If GF remains in the same range as the reported values, the metric is insensitive to the consolidation step and cannot support the claim that L2GTX's merging preserves faithfulness. A second, stronger observation: evaluate the global explanation's fidelity by holding out instances not used in the sampling and checking whether the global event summary predicts the black-box's decis

Watch

Extended reading notes

Core claim

The central claim is that a class-wise global explanation of a time-series classifier can be synthesised from local explanations without any access to the model's internals, and that this synthesis can be made compact without sacrificing faithfulness. L2GTX does this by (1) obtaining LOMATCE local explanations, in which each instance's decision-relevant behaviour is represented as clusters of parameterised event primitives (increasing/decreasing trends, local maxima/minima) with importance weights; (2) merging similar clusters across instances via agglomerative clustering controlled by a merge percentile; (3) computing global cluster importances and selecting a budgeted set of representative

Load-bearing premise

The evaluation rests on the assumption that the average fidelity of the local surrogate explanations over the selected instances (Global Faithfulness, Eq. 9) measures the quality of the global explanation, even though that average is computed from the local explanations alone and does not change meaningfully when clusters are merged or instances selected.

Editorial extensions

If this is right

  • L2GTX offers a model-agnostic route to class-level explanations for any time-series classifier, regardless of architecture (the paper demonstrates this on a convolutional and an LSTM-based network).
  • The merge percentile p provides a single tunable knob that trades compactness against detail: higher p yields fewer global clusters while the reported global faithfulness stays stable.
  • The produced explanations are expressed in human-readable terms (trends and extrema), which could help domain experts validate model behaviour in fields like healthcare and finance.
  • Because the method relies on a small sample of instances per class (15–30), it can be applied post-hoc to already trained classifiers without re-training or access to internal representations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's faithfulness metric (mean local surrogate R² over the selected instances) is computed from the local LOMATCE explanations and does not depend on the merging or instance-selection steps; stable global faithfulness across merge percentiles therefore cannot by itself distinguish a meaningful global summary from a random grouping of local events. A stricter test would measure how well the
  • A concrete extension is to compare L2GTX's summaries against baselines that merge clusters randomly or select instances greedily by importance without geometric merging; if global faithfulness remains indistinguishable, the metric is not capturing the consolidation step's contribution.
  • The interaction between the instance-selection budget B and the merge percentile p is underexplored: when B is small, the greedy coverage selection may be dominated by the first few instances, and it would be informative to report how the content and faithfulness of global explanations vary as B changes.
  • Since LOMATCE's local clusters are obtained by K-means with silhouette-based K, the global clusters inherit sensitivity to the local clustering hyperparameters; the paper does not report how variations in these hyperparameters affect the stability of the final global explanations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. L2GTX is a model-agnostic local-to-global explanation framework for time-series classifiers. It first generates local explanations with LOMATCE, then merges event-primitive clusters across instances, computes global cluster importance, greedily selects a budgeted subset of representative instances, and aggregates the events of covered clusters into class-wise summaries. The experiments on six UCR datasets with FCN and LSTM-FCN report 'Global Faithfulness' (GF) as the mean local surrogate R^2 over the selected instances, which is shown to remain stable across merge percentiles. The paper concludes that L2GTX yields compact, faithful class-wise global explanations.

Significance. Global explanation synthesis for time-series classifiers is an underexplored area, and the paper proposes a coherent, architecture-agnostic pipeline that maps local event-level explanations to class-level summaries. The use of parameterised event primitives and the explicit formulation of the clustering/selection/summarisation steps are strengths, and the method is, at a high level, reproducible. However, the core claim of 'faithful global explanations' is not supported by the reported evaluation: the primary metric measures local surrogate fidelity, not the fidelity of the aggregated global summary, and the description of class-wise selection is ambiguous. As a result, the paper's central contribution is currently unvalidated, although the method may be salvageable with a substantially re-designed evaluation.

major comments (3)
  1. [§4.4, Eq. (9)] Global Faithfulness is defined as the mean local surrogate fidelity F(x_i) over the selected instances, where F(x_i) is the R^2 of the LOMATCE ridge surrogate computed in Step 1. This quantity does not evaluate the global explanation produced by Steps 2–5: it is not affected by the cluster merging (Eq. 1), global importance (Eq. 3), or event summarisation (Eqs. 6–8), except indirectly through the composition of S. Thus the stability reported in Tables 3 and 4 is expected if the selected instances do not change drastically, and it would also be consistent with an arbitrarily bad global summary. The paper's own Section 4 introduction frames this as an 'expectation' rather than a validated property. The central claim requires a metric that directly assesses the global summary, for example by testing whether the summary statistics predict black-box outputs, or by comparing the summary agains
  2. [§3 (intro) and §4.3] The text describes the method as producing class-wise global explanations, but the algorithm pools N = C·ninst instances and then selects B instances globally from this pool (Step 4). Section 4.3 sets B equal to ninst. If B = ninst, the greedy, class-unconstrained selection can choose instances from a single class, contradicting the initial 'class-balanced' guarantee; if the pipeline is instead run separately for each class, then N should be ninst and the pooling notation is misleading. Because the experimental results are macro-averaged across classes, it is essential to specify precisely how class-wise summaries are derived and what budget is used per class.
  3. [§5] The quantitative evaluation lacks any baseline, ablation, or null model for the global aggregation itself. The only quantitative evidence is the GF tables and the cluster-count reduction in Fig. 2; the latter shows compaction but not fidelity or usefulness. The qualitative case studies (Figs. 3–6) cite known domain patterns, but they are not a substitute for a quantitative measure of global explanation quality. At minimum, the authors should compare against random instance selection, random merging, a global surrogate/rule-based baseline, and an oracle/upper bound. Without these, the aggregation step at the heart of L2GTX remains unvalidated.
minor comments (6)
  1. [Algorithm 1] Line 22 reuses the symbol S for both the selected instance set and the output summary; rename the summary to avoid confusion. The input line mentions 'Threshold τ', but the text defines τ via the merge percentile p; list p as an explicit input and clarify the relationship.
  2. [§4.3] The sentence 'B was set equal to ninst' is difficult to reconcile with N = C·ninst and the global selection in Step 4. Please clarify the budget for the pooled setting or state explicitly that selection is done per class.
  3. [§5, Tables 3–4] The claim of 'stable' faithfulness relies on overlapping confidence intervals; this is not a formal statistical test. Also, some values are rather low (e.g., GunPoint ~0.59, CBF LSTM ~0.51), and the paper does not discuss whether such values constitute 'faithful' explanations.
  4. [Figures 3–6] The caption says importance scores are normalised so that their sum equals one, but the paper does not specify the normalisation applied to the global importance scores from Eq. (3). Please state the normalisation.
  5. [§A.1] The LOMATCE description in the appendix omits important parameter values (e.g., number of perturbed samples N_z, bandwidth σ, top-n clusters per instance) and the software/version used. These details are needed for reproducibility.
  6. [General] Minor typographical and notation issues: 'ninst' formatting, 'K-means' vs 'k-means', and the statement 'class-balanced global explanations' could be read as balanced selection rather than balanced sampling. Please rephrase.

Circularity Check

1 steps flagged · score 6.0 of 10

Global Faithfulness (Eq. 9) is defined as mean local surrogate fidelity, so stable GF across merge percentiles is an artifact of the metric and does not validate the global explanation.

  1. self definitional [Section 4.4, Eq. (9); Section 3.1 (Step 1); Section 5, Fig. 2 caption]
    "The primary evaluation metric is Global Faithfulness (GF), which quantifies how accurately a global explanation summary reflects the behaviour of the underlying black-box model. Given a summary set S with budget B (i.e., |S|=B), GF is defined as the mean local fidelity, measured using the coefficient of determination (R2), across the selected instances: GF(S) = 1/|S| Σ_{xi∈S} F(xi), (9) where F(xi) denotes the local surrogate fidelity for instance xi."

    By Eq. (9), 'global faithfulness' is literally the mean of the local surrogate R2 values produced by LOMATCE in Step 1, before clustering, selection, or aggregation. The final global explanation is produced only in Step 5 by flattening clusters and computing attribute statistics (Eqs. 6–8); F(xi) does not depend on those summaries. Therefore the stable GF values in Tables 3–4, and the conclusion that consolidation 'does not degrade explanation faithfulness,' measure the local explainer's fit, not whether the aggregated global explanation reflects the black box. A global explanation could be arbitrarily wrong while GF stays high; the reported stability is forced by the metric's definition rather than by the quality of the global synthesis.

full rationale

The algorithmic pipeline itself—local explanation extraction, meta-clustering, instance selection, and event aggregation—is not internally circular: each step is defined on the outputs of the previous step and does not assume its own conclusion. The extensive use of LOMATCE is a self-citation, but it is ordinary reuse of a prior method and is not load-bearing in the sense of an unverified uniqueness claim; LOMATCE is the component being aggregated. The circularity is concentrated in the evaluation: Global Faithfulness (Eq. 9) is defined as the mean local surrogate fidelity of the very instances whose local explanations are the inputs to aggregation. This makes the headline result—stable GF across merge percentiles—an artifact of the metric, not evidence about the global explanation. No independent metric targets the global output (e.g., fidelity of the summarized event statistics to the black-box model), so the central claim that L2GTX yields faithful class-wise global explanations is only partially supported. This warrants a score of 6 rather than 0–2; it is not 8–10 because the method's clustering, selection, and summarization components have independent algorithmic content, and the circularity is confined to the validation metric.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities; it combines existing concepts (event primitives, local surrogates, hierarchical clustering, submodular selection). The main load-bearing assumptions are the validity of the local surrogate as a proxy for global faithfulness and the representativeness of greedy selection.

free parameters (3)
  • merge percentile p
    User-defined cut on hierarchical clustering merge distances; varied over {25,50,75,95} in experiments. The choice affects the number of global clusters and the resulting explanations.
  • instances per class ninst = 15 or 30
    User-selected number of instances sampled per class for explanation synthesis; set to 15 for small/medium datasets and 30 for larger ones.
  • instance selection budget B = equal to ninst
    Budget for the number of selected representative instances; set equal to ninst in experiments. Controls the size of the summarised explanation.
assumptions (5)
  • domain assumption Local surrogate fidelity R^2 is a valid measure of explanation faithfulness.
    Standard assumption in LIME-style local explanation literature, invoked in Section 4.4 when defining Global Faithfulness.
  • domain assumption Aggregating local explanations via importance-weighted cluster coverage preserves global faithfulness.
    This is the central hypothesis of L2GTX; the paper provides no direct test of whether the aggregated global summary reflects the model's class-level behaviour.
  • domain assumption Euclidean distance between cluster centroids captures semantic similarity of temporal event primitives.
    Used in Step 2 (Section 3.2) to merge local clusters; no justification beyond geometric proximity in parameter space.
  • domain assumption Greedy coverage maximisation selects a representative set of instances.
    Adapted from SP-LIME; the paper does not validate representativeness beyond coverage of top clusters.
  • domain assumption Mean and standard deviation of event attributes provide a concise and accurate summary of a cluster.
    Used in Step 5 (Section 3.5) to produce the final global explanation; assumes events within a cluster are sufficiently homogeneous.

how reviews work

0 comments
Cite this review

Pith. "Pith review of L2GTX: From Local to Global Time Series Explanations." pith.science (2026). https://pith.science/paper/WGOAESF3

@misc{pith2026260313065,
  author       = {Pith},
  title        = {Pith review of: L2GTX: From Local to Global Time Series Explanations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WGOAESF3}},
  note         = {Machine review of arXiv:2603.13065}
}
read the original abstract

Deep learning models achieve high accuracy in time series classification, yet understanding their class-level decision behaviour remains challenging. Explanations for time series must respect temporal dependencies and identify patterns that recur across instances. Existing approaches face three limitations: model-agnostic XAI methods developed for images and tabular data do not readily extend to time series, global explanation synthesis for time series remains underexplored, and most existing global approaches are model-specific. We propose L2GTX, a model-agnostic framework that generates class-wise global explanations by aggregating local explanations from a representative set of instances. L2GTX extracts clusters of parameterised temporal event primitives, such as increasing or decreasing trends and local extrema, together with their importance scores from instance-level explanations produced by LOMATCE. These clusters are merged across instances to reduce redundancy, and an instance-cluster importance matrix is used to estimate global relevance. Under a user-defined instance selection budget, L2GTX selects representative instances that maximise coverage of influential clusters. Events from the selected instances are then aggregated into concise class-wise global explanations. Experiments on six benchmark time series datasets show that L2GTX produces compact and interpretable global explanations while maintaining stable global faithfulness measured as mean local surrogate fidelity.

Figures

Figures reproduced from arXiv: 2603.13065 by the authors.

Figure 1
Figure 1. Overview of the L2GTX method. (1) Local explanations are generated for a batch of time-series instances [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Effect of the agglomerative clustering threshold (merge percentile) on the number of class-averaged global [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Class-wise global explanations produced by L2GTX for the ECG200 dataset using the FCN model. Bars [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Class-wise global explanations produced by L2GTX for the ECG200 dataset using the LSTM-FCN model. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Class-wise global explanations produced by L2GTX for the Coffee dataset using the FCN model at merge [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Class-wise global explanations produced by L2GTX for the Coffee dataset using the LSTM-FCN model at [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 5 linked inside Pith

  1. [1]

    Financial series prediction using attention lstm.arXiv preprint arXiv:1902.10877,

    Sangyeon Kim and Myungjoo Kang. Financial series prediction using attention lstm.arXiv preprint arXiv:1902.10877,

    arXiv 1902

  2. [5]

    Interpretation of deep temporal representations by selective visualization of internally activated nodes.arXiv preprint arXiv:2004.12538,

    Sohee Cho, Ginkyeng Lee, Wonjoon Chang, and Jaesik Choi. Interpretation of deep temporal representations by selective visualization of internally activated nodes.arXiv preprint arXiv:2004.12538,

    arXiv 2004

  3. [7]

    Global aggregations of local explanations for black box models.arXiv preprint arXiv:1907.03039,

    Ilse Van Der Linden, Hinda Haned, and Evangelos Kanoulas. Global aggregations of local explanations for black box models.arXiv preprint arXiv:1907.03039,

    arXiv 1907

  4. [1999]

    Interpreting black-box time series classifiers using parameterised event primitives.xAI-2024 Late-breaking Work, Demos & Doctoral Consortium Joint Proceedings, 2024a

    Ephrem T Mekonnen, Luca Longo, and Pierpaolo Dondio. Interpreting black-box time series classifiers using parameterised event primitives.xAI-2024 Late-breaking Work, Demos & Doctoral Consortium Joint Proceedings, 2024a. 15 L2GTX: From Local to Global Time Series Explanations Christoph Molnar.Interpretable Machine Learning. Christoph Molnar, 3 edition,

    2024

  5. [2019]

    Explainable artificial intelligence (xai) on timeseries data: A survey.arXiv preprint arXiv:2104.00950,

    Thomas Rojat, Raphaël Puget, David Filliat, Javier Del Ser, Rodolphe Gelin, and Natalia Díaz-Rodríguez. Explainable artificial intelligence (xai) on timeseries data: A survey.arXiv preprint arXiv:2104.00950,

    arXiv

  6. [2020]

    A global model-agnostic rule-based xai method based on parameterised event primitives for time series classifiers.Frontiers in Artificial Intelligence, 7:1381921, 2024b

    Ephrem T Mekonnen, Pierpaolo Dondio, and Luca Longo. A global model-agnostic rule-based xai method based on parameterised event primitives for time series classifiers.Frontiers in Artificial Intelligence, 7:1381921, 2024b. Ephrem Mekonnen, Pierpaolo Dondio, and Luca Longo. Explaining deep learning time series classification models using a decision tree-ba...

    2023

  7. [2021]

    Explainable ai for time series via virtual inspection layers.arXiv preprint arXiv:2303.06365,

    Johanna Vielhaben, Sebastian Lapuschkin, Grégoire Montavon, and Wojciech Samek. Explainable ai for time series via virtual inspection layers.arXiv preprint arXiv:2303.06365,

    arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.