L2GTX: From Local to Global Time Series Explanations
Authors: Ephrem Tibebe Mekonnen, Luca Longo, Lucas Rizzo, Pierpaolo Dondio
Organizations: School of Computer Science, Technological University Dublin Dublin, Ireland · Artificial Intelligence and Cognitive Load Research Lab, University College Cork Cork, Ireland · School of Computer Science and Information Technology, University College Cork Cork, Ireland · School of Computing, Dublin City University Dublin, Ireland
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 & tables
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: Illustrative examples of local explanations generated by LOMATCE for the FCN classifier on the ECG200 dataset.
Name
Data Size
No. Classes
Length
Type
ECG200
200
2
96
Medical
GunPoint
200
2
150
Action
Coffee
48
2
286
Food
FordA
4921
2
500
Sensor
FordB
4446
2
500
Sensor
CBF
900
3
128
Synthetic
Table 1: Descriptive statistics of the six UCR datasets used in the experiment.
Dataset
FCN
LSTM-FCN
Test
Val
Test
Val
ECG200
0.87 ± 0.03
0.86 ± 0.04
0.86 ± 0.01
0.85 ± 0.01
GunPoint
0.99 ± 0.01
0.98 ± 0.02
0.98 ± 0.01
0.98 ± 0.01
Coffee
1.00 ± 0.00
1.00 ± 0.00
1.00 ± 0.00
1.00 ± 0.00
FordA
0.90 ± 0.02
0.90 ± 0.02
0.88 ± 0.01
0.87 ± 0.01
FordB
0.88 ± 0.02
0.89 ± 0.02
0.86 ± 0.01
0.86 ± 0.01
Table 2: Mean test and validation accuracy (Test Acc and Valid Acc) for each dataset–model pairs, reported as mean ± 95% confidence interval.
Dataset
p = 25
p = 50
p = 75
p = 95
ECG200
0.784 ± 0.015
0.788 ± 0.013
0.780 ± 0.026
0.792 ± 0.014
GunPoint
0.593 ± 0.007
0.599 ± 0.019
0.601 ± 0.007
0.597 ± 0.011
Coffee
0.683 ± 0.010
0.678 ± 0.006
0.678 ± 0.005
0.678 ± 0.015
FordA
0.674 ± 0.021
0.672 ± 0.029
0.673 ± 0.021
0.672 ± 0.028
FordB
0.675 ± 0.008
0.679 ± 0.034
0.673 ± 0.006
0.673 ± 0.029
CBF
0.625 ± 0.018
0.626 ± 0.011
0.633 ± 0.016
0.625 ± 0.008
Table 3: Global faithfulness (mean ± 95% CI) of L2GTX explanations using the FCN classifier under different merge percentiles. Results are macro-averaged across classes and aggregated over three random seeds.
Dataset
p = 25
p = 50
p = 75
p = 95
ECG200
0.828 ± 0.010
0.832 ± 0.013
0.829 ± 0.021
0.831 ± 0.007
GunPoint
0.617 ± 0.074
0.619 ± 0.067
0.588 ± 0.086
0.638 ± 0.011
Coffee
0.617 ± 0.008
0.609 ± 0.004
0.616 ± 0.036
0.608 ± 0.003
FordA
0.618 ± 0.028
0.621 ± 0.015
0.614 ± 0.039
0.627 ± 0.035
FordB
0.661 ± 0.021
0.656 ± 0.039
0.651 ± 0.050
0.655 ± 0.027
CBF
0.519 ± 0.020
0.508 ± 0.025
0.519 ± 0.033
0.502 ± 0.015
Table 4: Global faithfulness (mean ± 95% CI) of L2GTX explanations using the LSTM-FCN classifier under different merge percentiles. Results are macro-averaged across classes and aggregated over three random seeds.
Figure 3: 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 4: 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 5: 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 6: 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.
Figure 7: Class-wise global explanations produced by L2GTX for the Coffee dataset using the LSTM-FCN model at merge percentile 95.
Technische Hochschule Mittelhessen - University of Applied Science, Friedberg, Germany · Hessian Center for AI (hessian.AI), Darmstadt, Germany · Marburg University, Marburg, Germany