Evaluation of Interpretability Methods for Time-Series Classification Models

Hugues Turbé from the University of Geneva provides datasets, trained models, and results for evaluating post-hoc interpretability methods. The data includes a processed ECG dataset from a Cardiology Challenge, a time-series dataset from the UCR Archive, and a synthetic dataset created for the study. The repository contains relevance scores and evaluation metrics for models including CNN, transformer, and bi-LSTM architectures.

Use Cases

• Benchmarking post-hoc interpretability methods based on relevance scores and metrics provided.
• Training and testing time-series classifiers on a processed ECG dataset.
• Evaluating model performance on a synthetic dataset developed for controlled interpretability analysis.
• Comparing the interpretability of CNN, transformer, and bi-LSTM architectures on common time-series tasks.

Strengths

• Includes three distinct datasets: a processed medical ECG dataset, a standard UCR archive dataset, and a purpose-built synthetic dataset.
• Provides pre-trained models (CNN, transformer, bi-LSTM) and computed relevance metrics for multiple interpretability methods.
• Results are summarized in an Excel file for cross-dataset comparison.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file sizes are unknown, which may limit suitability assessment.

Provenance

Source

University of Geneva

Collection Method

Combines a processed version of the CPSC ECG dataset, a dataset from the UCR Time Series Classification Archive, and a synthetic dataset created for the study.