L2GTX: From Local to Global Time Series Explanations
| Authors | Ephrem Tibebe Mekonnen et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2603.13065 |
| Download | |
| Categories | cs.LG, cs.AI |
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.
Engineering Breakdown
Plain English
L2GTX solves a critical problem in interpretable machine learning: explaining why deep learning models make class-level decisions on time series data. Existing explanation methods (like LIME and SHAP) work well for images and tabular data but fail on time series because they don't respect temporal dependencies and don't identify recurring patterns across multiple instances. The authors propose L2GTX, a model-agnostic framework that aggregates local explanations from representative instances to generate global, class-wise explanations by extracting clusters of temporal event primitives like "increasing" or "decreasing" trends. This approach enables practitioners to understand what patterns a time series classifier learned for each output class without needing access to model internals.
Core Technical Contribution
L2GTX introduces a novel two-stage aggregation pipeline that bridges local and global explanation for time series classification. The core innovation is extracting parameterized temporal event primitives—discrete, interpretable building blocks like monotonic segments, peaks, and valleys—from local explanations and clustering them to identify class-level patterns. Unlike prior global explanation methods that are typically model-specific (tied to RNNs, CNNs, or Transformers), L2GTX is genuinely model-agnostic because it only requires access to model predictions and local explanations, not internal weights or gradients. This is technically novel because previous attempts to globalize local explanations either required architectural knowledge or produced outputs that didn't respect temporal structure.
How It Works
L2GTX operates in three distinct phases. First, for each instance in the dataset, it generates a local explanation (e.g., using gradient-based or perturbation-based methods) that identifies which time steps or temporal regions influenced the prediction. Second, these local explanations are converted into a standardized representation: temporal event primitives with parameters like start time, duration, and magnitude of change. Third, the framework clusters these primitives across instances assigned to the same class using unsupervised learning (e.g., k-means or hierarchical clustering), then aggregates representative primitives into class-level summaries. The output is a set of interpretable, recurring temporal patterns per class—for instance, "all instances classified as 'anomaly' exhibit a 200ms spike followed by sustained elevation for 5 seconds." This design respects temporal dependencies by anchoring explanations to time-aware features rather than treating time series as unordered feature vectors.
Production Impact
For engineers deploying time series classifiers in domains like healthcare monitoring, industrial predictive maintenance, or financial fraud detection, L2GTX immediately enables model auditing and debugging without retraining or architectural changes. When a model makes unexpected predictions on a new class of data, practitioners can now visualize what temporal patterns the model learned for that class, allowing rapid identification of data drift, label errors, or spurious correlations. Integration is straightforward: feed predictions through any local explanation method, then run the L2GTX aggregation pipeline (likely a few hundred lines of Python); no model retraining required. The trade-off is computational: generating local explanations for all training instances can be expensive (linear in dataset size), and clustering primitives requires setting hyperparameters like the number of clusters per class, which adds tuning overhead. In latency-sensitive applications, storing and querying global explanations is fast (microseconds), but the upfront analysis may take minutes to hours depending on dataset size.
Limitations and When Not to Use This
L2GTX assumes that local explanations are reliable and semantically meaningful, which breaks down if the underlying explanation method produces noise or artifacts—garbage in, garbage out. The framework requires choosing how to parameterize temporal primitives and which clustering algorithm to use, introducing hyperparameter sensitivity that the paper doesn't fully characterize; there's no principled way to determine the "right" number of clusters per class. The approach also assumes explanations are aggregatable across instances, which may fail when temporal patterns are rare or highly instance-specific—in that case, global summaries become too coarse or miss important edge cases. The paper abstract doesn't mention evaluation on very long time series (e.g., hours or days of high-frequency data) or multivariate time series with dependencies between channels, so scalability and applicability to these domains remains unclear.
Research Context
L2GTX builds on two mature research threads: local explanation methods (LIME, SHAP, gradient-based approaches) originally developed for computer vision and tabular data, and the emerging push for global explanations in neural networks (pioneered by concept activation vectors and prototype-based interpretability). The work directly addresses a gap identified by recent surveys on time series XAI, which noted that most methods either extract local explanations poorly or attempt global explanations in a model-specific manner (e.g., attention visualization for Transformers). By demonstrating a model-agnostic aggregation strategy specifically designed for temporal structure, this paper advances the broader interpretability agenda toward practical, modality-aware explanation techniques. Future work likely extends this to other sequential modalities (NLP, video) and explores whether global explanations improve downstream tasks like data augmentation or transfer learning.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::