L2GTX: From Local to Global Time Series Explanations

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 (6)

• Figure 1: Overview of the L2GTX method. (1) Local explanations are generated for a batch of time-series instances using LOMATCE. (2) Similar event clusters are merged to reduce redundancy, producing an instance–cluster importance matrix. (3) Global importance scores are computed for the clusters. (4) Given an instance selection budget $B$, the instance–cluster importance matrix, and the global cluster importance scores, a representative subset of instances that maximises the coverage of influential clusters is selected. (5) Parameterised event primitives (PEPs) from the selected instances are aggregated to produce the final class-wise global explanation.
• Figure 2: Effect of the agglomerative clustering threshold (merge percentile) on the number of class-averaged global clusters. Increasing the merge percentile progressively consolidates event clusters, resulting in a monotonic reduction in the total number of global clusters. Across both models, this consolidation does not degrade explanation faithfulness, as reflected by stable or improved class-averaged global faithfulness scores.
• Figure 3: Class-wise global explanations produced by L2GTX for the ECG200 dataset using the FCN model. Bars indicate the global importance of aggregated temporal event clusters, with colours denoting event types.
• Figure 4: Class-wise global explanations produced by L2GTX for the ECG200 dataset using the LSTM-FCN model. Bars indicate the global importance of aggregated temporal event clusters, with colours denoting event types.
• Figure 5: Class-wise global explanations produced by L2GTX for the Coffee dataset using the FCN model at merge percentile 95. Bars denote the global importance of aggregated temporal event clusters, with colours indicating event dynamics.