PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards A Unified Information Bottleneck Framework for Time Series Explanations

Xu Zheng, Zichuan Liu, Zhuomin Chen, Mayur Akewar, Janki Bhimani, Jason Liu, Mo Sha, Jingchao Ni, Wei Cheng, Dongsheng Luo

arXiv:2608.25897v1cs.LGcs.AI

TL;DR

Time-series explainability remains divided between attribution and counterfactual methods, with the former lacking causal validation and the latter producing unstable, adversarial-like changes. TimeX++ unifies both through an Information Bottleneck objective, generating in-distribution attributions and stable counterfactuals, and reports state-of-the-art performance across synthetic and real-world datasets.

  • Problem

    Attribution and counterfactual explanations are studied as disconnected tasks, leaving attributions without causal verification and counterfactuals without semantic grounding.

  • Method

    TimeX++ uses an Information Bottleneck-based unified objective to generate in-distribution attribution instances and semantically grounded counterfactuals.

  • Results

    TimeX++ achieves state-of-the-art performance for both attribution and counterfactual explanation tasks on synthetic and real-world datasets.

  • Takeaways & Limitations

    The framework provides faithful attributions and robust counterfactual explanations while reflecting complex behaviors of pretrained time-series classifiers.

Abstract

from arXiv · show

Explaining deep learning models operating on time series data is crucial in various applications that require transparent and interpretable insights into model behavior. {Existing explanation methods generally fall into two categories: attribution-based explanations, which identify the temporal regions most responsible for a prediction, and counterfactual explanations, which reveal how an input should be modified to alter the model's decision.} {Despite valuable insights, these two fields are largely studied independently. This disconnect leaves attribution methods lacking causal validation, while counterfactual methods suffer from severe instability, producing adversarial-like noise instead of meaningful explanations.} In this work, we revisit time-series explainability from an information-theoretic perspective and show that existing explainers are vulnerable to trivial solutions and distributional shifts. To address these limitations, we propose a unified objective function for explainable time series learning that bridges attribution and counterfactual reasoning within a single framework. Building upon the Information Bottleneck principle, our formulation explicitly prevents trivial explanations and out-of-distribution counterfactuals. {Based on this objective function, we introduce {\modelname}, a novel explanation framework that learns a parametric transformation network to construct explanation-embedded instances, where preserved information yields attribution explanations and controlled information removal produces stable counterfactual explanations.} We evaluate {\modelname} on synthetic and real-world benchmarks against state-of-the-art baselines. Extensive quantitative and qualitative results show that {\modelname} consistently outperforms competing methods, yielding faithful attributions and stable counterfactual explanations.

I. INTRODUCTION

Time-series explainability is split between attribution and counterfactual methods, leaving causal validation and semantic stability insufficiently addressed. The paper unifies both through an Information Bottleneck objective and proposes TimeX++ to produce in-distribution attributions and stable counterfactuals.

  • Motivation: Attribution methods identify salient temporal regions, but commonly use ad-hoc objectives and can produce out-of-distribution samples during evaluation.Masking or substituting salient features may violate the data manifold, making classifier-based evaluations unreliable.
  • Motivation: Treating attribution and counterfactual explanations separately leaves attributions without causal verification and counterfactuals vulnerable to adversarial noise.Counterfactuals lack semantic grounding when label-flip searches modify the entire time series without attribution guidance.
  • Unified framework: The Information Bottleneck principle connects explanations by balancing compactness, I(X; X′), against informativeness, I(X′; Y ).Attribution is framed as retaining label-sufficient information while discarding redundant temporal dynamics.
  • Unified framework: TimeX++ generates in-distribution explanation-embedded instances for attribution and constrains counterfactual perturbations to meaningful feature dimensions.The framework uses an attribution bottleneck to support semantically grounded counterfactual generation.
  • Results: TimeX++ achieves state-of-the-art performance on attribution and counterfactual tasks across synthetic and real-world datasets.The paper evaluates both explanation types against existing methods and reports their effectiveness.

III. PROBLEM FORMULATION

The paper formulates time-series attribution and counterfactual explanation as related information-bottleneck problems. It defines their objectives, identifies computational and distributional challenges, and replaces intractable mutual-information terms with tractable constraints.

  • III. PROBLEM FORMULATION: The task concerns post-hoc, instance-level explanations for a black-box time-series classifier operating on multivariate or univariate inputs.Inputs are sequences X ∈ R^T×D with labels from a finite label space.
  • III. PROBLEM FORMULATION: Attribution seeks a masked sub-instance X′ = X ⊙ M that explains the classifier’s prediction.The mask M is binary over temporal positions and features.
  • III. PROBLEM FORMULATION: A differentiable extractor maps X to mask probabilities and samples a mask through independent Bernoulli variables.The extractor produces M ∈ [0,1]^T×D probabilities before discrete sampling.
  • III. PROBLEM FORMULATION: Counterfactual explanation seeks a minimally perturbed X′′ classified as a target label, with perturbations restricted to the attribution mask’s semantic support.The objective requires f(X′′) = Y′′ while minimizing d(X, X′′).
  • IV. EXPLAINING TIME SERIES LEARNING VIA INFORMATION BOTTLENECK: The Information Bottleneck view balances label informativeness against compactness by optimizing I(X′; Y) − αI(X; X′).The paper replaces direct informativeness estimation with a tractable label-consistency measure.
  • IV. EXPLAINING TIME SERIES LEARNING VIA INFORMATION BOTTLENECK: Direct application is challenged by intractable mutual-information estimation, out-of-distribution attribution inputs, and adversarially unstable counterfactuals.Masked instances can violate the data manifold, while unconstrained counterfactual optimization can exploit classifier vulnerabilities.
  • A. Tractable Compactness Bound: The compactness term is replaced by a variational upper bound that penalizes mask size and entropy through a Bernoulli prior.This yields a KL-based tractable constraint that compresses the bottleneck without complex mutual-information estimators.

B. Distribution-Preserving and Bounded Generation

TimeX++ evaluates masked explanations through generated in-distribution instances and constrains counterfactual changes to the shared attribution bottleneck. Its unified objective combines compactness, label consistency, distribution preservation, and structural bounding.

  • B. Distribution-Preserving and Bounded Generation: The framework projects the bottleneck into an in-distribution generated instance before evaluating the classifier.This design addresses attribution’s OOD evaluation problem and counterfactual instability within one formulation.
  • B. Distribution-Preserving and Bounded Generation: Distribution preservation regularizes the generated instance toward the true data manifold through a KL penalty.The penalty is LKL(P_X̃∥P_X).
  • B. Distribution-Preserving and Bounded Generation: The final objective combines the tractable compactness bound with label consistency, distribution preservation, and a bottleneck constraint.These terms are combined into a single unified optimization objective.
  • B. Distribution-Preserving and Bounded Generation: For attribution, the objective label equals the original label and the mask is minimized while generated evaluations remain in-distribution.The KL term is used to resolve the OOD problem during attribution.
  • B. Distribution-Preserving and Bounded Generation: For counterfactuals, the target label is used while the bottleneck constraint keeps the trajectory within the mask’s semantic support.This restricts manipulation to features identified by attribution.
  • B. Distribution-Preserving and Bounded Generation: TimeX++ implements the objective with a shared extractor and paradigm-specific conditioned generators.The configuration produces TimeXa++ and TimeXcf++ modes.
  • B. Distribution-Preserving and Bounded Generation: The extractor uses a Transformer to produce mask probabilities and a straight-through estimator to sample binary masks.The sampled mask is used by the bottleneck formulation.
  • A. Attribution Bottleneck Extractor: Bottleneck extraction uses a sparse prior and a temporal continuity penalty to encourage compact, contiguous masks.The total extraction loss is LM = Lcompact + λconLcon.

B. Conditioned Generators

The conditioned generators transform attribution masks into explanation-embedded instances and generate counterfactual perturbations confined to masked regions. Separate generator designs address OOD attribution artifacts and adversarial counterfactual noise.

  • B. Conditioned Generators: Conditioned generators translate the structural bottleneck M into generated explanation-embedded instances in the data space.ψa targets attribution OOD issues, while ψcf targets counterfactual robustness.
  • B. Conditioned Generators: TimeXa++ combines the masked input with a dataset-sampled background and uses ψa to generate a smoothed instance.Distribution preservation and a reference-distance constraint regulate the generated attribution instance.
  • B. Conditioned Generators: TimeXcf++ confines counterfactual perturbations to the semantic bottleneck to avoid exploiting classifier null spaces.The method contrasts with unconstrained perturbation in continuous temporal space.
  • B. Conditioned Generators: Training computes compactness, reference, distribution, and label-consistency losses before updating the extractor and active generators.The pipeline alternates generator-specific optimization according to the selected explanation mode.
  • B. Conditioned Generators: The counterfactual generator produces a target-driven perturbation matrix E from X and M.The structural form uses the input, mask, perturbation matrix, and auxiliary variation.
  • B. Conditioned Generators: A variation function incorporates dynamics from a training instance with the target label, while inference removes the auxiliary noise term.During inference, the counterfactual simplifies to X̃cf = X + M ⊙ E.
  • B. Conditioned Generators: The bottleneck constraint penalizes deviations from the original reference in non-bottleneck background regions.This acts as a hard anchor against imperceptible adversarial changes in redundant features.

C. End-to-End Optimization

TimeX++ shares one architecture and mathematical objective across attribution and counterfactual modes. Switching the objective label and active generator enables end-to-end optimization for either explanation type.

  • C. End-to-End Optimization: The entire framework is optimized end-to-end using a shared architecture and unified objective.Both TimeXa++ and TimeXcf++ minimize the same overarching formulation.
  • C. End-to-End Optimization: Attribution and counterfactual training alternate by toggling the objective label and selecting the corresponding generator.The procedure is summarized in Algorithm 1.

VI. EXPERIMENTS

The experiments evaluate TimeX++ across attribution and counterfactual explanation tasks using diverse datasets, baselines, and quantitative metrics.

  • The evaluation targets faithful attributions, effective counterfactuals, reduced out-of-distribution shift, and improved inference efficiency.
  • The experimental setup combines diverse datasets, established baselines, and strict evaluation metrics.

1) Datasets:

The study uses ten synthetic and real-world time-series datasets, compares eleven explainers, and evaluates attribution and counterfactual quality with complementary metrics.

  • Datasets: Ten datasets span synthetic benchmarks with ground-truth explanations and real-world scenarios across multiple application domains.
  • Baselines: TimeX++ is benchmarked against eleven state-of-the-art explainers spanning attribution and counterfactual paradigms.
  • Attribution Metrics: Attribution quality is assessed using ground-truth salient features with AUP, AUR, and AUPRC metrics.
  • Attribution Metrics: Precision and recall curves are defined from truly salient and selected index sets across thresholds.
  • Counterfactual Metrics: Counterfactuals are evaluated by validity, confidence, sparsity, Proximity-L1, and Proximity-L2, with direction-specific preferences.

B. Efficacy of Unified Explanations

TimeX++ is evaluated through a unified training process for attribution and counterfactual explanations, with strong attribution results and Pareto-based counterfactual comparison.

  • Unified Evaluation: The unified training process evaluates TimeX++ jointly on attribution fidelity and counterfactual efficacy.
  • Attribution: TimeXa++ outperforms competing explainers in 9 of 12 synthetic attribution cases.
  • Attribution: 11.01% average AUPRC, 10.87% AUP, and 1.25% AUR improvements over TIMEX are reported across synthetic attribution evaluations.
  • Attribution: On real-world occlusion experiments, TimeXa++ outperforms other explainers on Epilepsy, PAM, and Boiler datasets.
  • Counterfactuals: Targeted counterfactual explanations are compared across all target labels using Pareto frontier analysis because validity, confidence, sparsity, and proximity conflict.
  • Counterfactuals: Untargeted experiments on four synthetic datasets show trends consistent with the targeted setting, while CONFETTI is sensitive to reference-data selection.

C. Resilience to Out-of-Distribution Shift

The resilience analysis addresses distributional shift in attribution and adversarial degeneration in counterfactual explanations through distributional metrics, localized perturbations, and noise testing.

  • Distributional Alignment: KDE, KL divergence, and MMD quantify alignment between explanation-embedded instances and original temporal sequences.
  • Distributional Alignment: TimeXa++ explanations have significantly lower MMD and KL divergence than conventional masking strategies.
  • Counterfactual Robustness: TimeXcf++ confines perturbations to critical structural regions, producing localized changes rather than distributed adversarial noise.
  • Counterfactual Robustness: Under Gaussian noise injection, TimeXcf++ remains more robust than baselines on FreqShape and LowVar datasets.

E. Runtime Results

TimeX++ is evaluated for runtime across attribution and counterfactual explanation tasks, with efficient training and especially fast inference. Additional experiments indicate strong performance across classifier architectures and counterfactual objectives.

  • Training each experiment takes approximately 3 to 15 minutes per fold, depending on dataset volume.
  • TimeXa++ requires less overall attribution training time than TIMEX because it directly perturbs the black-box instead of training a white-box consistency model.
  • TimeX++ is the most expedient model during inference across both attribution and counterfactual tasks.
  • TimeXcf++ is over 100 times faster than the fastest counterfactual baseline, AB-CF, by using a single forward inference pass.
  • Under CNN predictors, TimeXa++ achieves the highest AUPRC and AUP on SeqComb-MV and ECG datasets.
  • Across CNN and LSTM backbones, TimeXcf++ balances validity, confidence, sparsity, and proximity, indicating generalizability across architectures.

G. Ablation Study and Hyperparameter Analysis

Ablation and hyperparameter analyses show that TimeX++ depends on discrete masking, distributional and structural losses, and careful control of sparsity. Visualizations further indicate faithful attribution regions and compact, pattern-related counterfactual modifications.

  • 39.78% higher AUPRC, 22.04% higher AUP, and 3.10% higher AUR are achieved on ECG when TimeXa++ uses STE against leading baselines.Across datasets, STE also increases AUPRC by an average of 8.87% versus continuous masking.
  • Removing LKL widens the distributional gap and reduces attribution performance, while removing Ldr causes the explainer to fail.
  • Removing LLC dramatically degrades counterfactual validity because the generator loses its primary signal for targeted label flipping.
  • Removing Ldr substantially worsens counterfactual proximity, while removing Lbound deteriorates sparsity and proximity by scattering perturbations across the sequence.
  • Attribution performance stabilizes for r between 0.4 and 0.7, whereas increasing r improves counterfactual validity and confidence but worsens sparsity and proximity.
  • On FreqShapes, TimeXa++ focuses on peaks matching ground-truth explanations, while TimeXcf++ makes fewer, pattern-related modifications than competing counterfactual methods.IG highlights unnecessarily large regions, TIMEX is affected by distribution bias, and several counterfactual baselines modify large input portions or fail validity.
  • The conclusion identifies dataset-dependent hyperparameter control over explanation quantifiers as a key limitation.
Loading 2608.25897v1…