New Algorithm Reveals How Deep Learning Models Classify Time Series Data
Chemistry

New Algorithm Reveals How Deep Learning Models Classify Time Series Data

Researchers introduce L2GTX, a model-agnostic method that bridges local and global explainability for sequential data without inspecting neural network internals.

By Bilal Abbasi
Published:
Email this Article
New Algorithm Reveals How Deep Learning Models Classify Time Series Data
Photo by Brecht Corbeel on Unsplash

Deep learning models achieve high accuracy in time series classification, yet understanding their class-level decision behaviour remains challenging. A new algorithm called L2GTX successfully extracts interpretable global explanations from black-box time series classifiers without inspecting their internal architecture.

Time series models are widely used in sensitive applications such as electrocardiogram monitoring for healthcare and sensor analysis in manufacturing. Despite their strong predictive performance, they operate as opaque black boxes that yield classifications without explaining which underlying temporal patterns drove the decision. This lack of transparency undermines trust, complicates error analysis, and creates barriers in regulated domains requiring strict accountability.

Overcoming Limitations in Sequence Explainability

While explainable artificial intelligence has developed mature techniques for static data types like images and tables, extending those methods to sequential data is difficult. Time series feature strong temporal dependencies, variable event durations, and shifting pattern locations across samples. Consequently, most existing explainability methods focus on local instance-level explanations, while global explanation approaches remain rare and tied to specific network architectures.

The research team developed L2GTX (Local-to-Global Time-series eXplanations) to bridge local and global interpretability. As published in Explainable Artificial Intelligence, the method operates in five sequential steps. It first generates local explanations for a sampled batch of instances using LOMATCE (LOcal Model-Agnostic Time Series Classification Explanations), which identifies parameterised temporal event primitives such as increasing or decreasing trends and local extrema.

Merging Patterns and Budgeted Instance Selection

To reduce redundancy, L2GTX merges similar event clusters across instances using agglomerative hierarchical clustering based on Euclidean distances between centroids. A user-defined merge percentile controls the level of consolidation, where higher percentiles yield fewer and more general global clusters. An instance-cluster importance matrix is then constructed to estimate the global relevance of each temporal pattern across the dataset.

Using a user-specified budget, the framework applies greedy submodular optimisation to select a diverse subset of representative instances that jointly maximise the coverage of influential clusters. The event attributes from the selected instances are flattened and summarised by computing their mean and standard deviation per primitive type, yielding concise class-wise global explanations.

Faithful Performance Across Diverse Benchmarks

The authors evaluated L2GTX on six univariate time series datasets from the UCR Archive, including medical, action, food, sensor, and synthetic domains. They tested the method using two distinct deep learning architectures, a Fully Convolutional Network and an LSTM-FCN combining convolutional layers with long short-term memory.

Quantitative evaluations measured global faithfulness via mean local surrogate fidelity using the coefficient of determination across different merge percentiles. The results demonstrated that L2GTX maintains stable global faithfulness even under aggressive explanation consolidation. Case studies on ECG and spectral coffee datasets confirmed that the synthesized global explanations align with domain-specific signal morphology while remaining sensitive to architectural differences.

Future work will extend the framework to multivariate time series to handle cross-channel interactions, alongside human-centred evaluations with domain experts to assess the practical utility of the generated summaries in real-world settings.

The research was published in Explainable Artificial Intelligence on November 19, 2026.

Fact Checked

This article has been fact checked for accuracy, with information verified against reputable sources. Learn more about us and our editorial process.

Last reviewed on .

Article history

  • Latest version

Reference(s)

  1. Mekonnen, Ephrem Tibebe., et al. “L2GTX: From Local to Global Time Series Explanations.” Communications in Computer and Information Science, November 19, 2026, pp. 27-48. Springer Nature Switzerland, doi: 10.1007/978-3-032-45470-6_2. <https://doi.org/10.1007/978-3-032-45470-6_2>.

Cite this page:

Abbasi, Bilal. “New Algorithm Reveals How Deep Learning Models Classify Time Series Data.” BioScience. BioScience ISSN 2521-5760, 19 November 2026. <https://www.bioscience.com.pk/en/subject/chemistry/l2gtx-from-local-to-global-time-series-explanations>. Abbasi, B. (2026, November 19). “New Algorithm Reveals How Deep Learning Models Classify Time Series Data.” BioScience. ISSN 2521-5760. Retrieved November 19, 2026 from https://www.bioscience.com.pk/en/subject/chemistry/l2gtx-from-local-to-global-time-series-explanations Abbasi, Bilal. “New Algorithm Reveals How Deep Learning Models Classify Time Series Data.” BioScience. ISSN 2521-5760. https://www.bioscience.com.pk/en/subject/chemistry/l2gtx-from-local-to-global-time-series-explanations (accessed November 19, 2026).
End of the article