Glacier: Guided Counterfactual Explanations
Glacier is a model-agnostic method for generating locally constrained counterfactual explanations in time series classification. It utilizes gradient search on either the original time series data or in a learned latent space, allowing researchers flexibility to apply constraints that promote realistic and context-sensitive modifications.
This approach is particularly suited for binary classification tasks with time series data, using constraints that guide changes to specific data segments while discouraging modifications to others. Glacier's flexibility extends to four configurations:
- Unconstrained: Applies changes across the entire time series.
- Example-specific: Focuses on specific time points in each example using LIMESegment.
- Global: Prioritizes important segments across all test examples.
- Uniform: Minimizes changes across the series for a broader application.
Glacier is compatible with modern classifiers and has been tested across 40 datasets from the UCR archive, supporting critical insights into model decisions by visualizing counterfactual impacts on time series classifications.
Wildboar: Shapelet-Based Time Series Classification
Wildboar is a shapelet-based method for time series classification that leverages shapelet transforms to identify significant patterns within the time series data. By using an efficient search strategy, Wildboar can extract shapelets (subsequences) that differentiate between classes, supporting both interpretability and high classification performance.
Wildboar is particularly suited for binary classification tasks where key patterns in the series can be leveraged to distinguish classes. With options to control the number of shapelets and depth of classification trees, Wildboar offers a highly customizable approach to time series counterfactuals. Key highlights include:
- Efficiency: Optimized shapelet discovery and tree-based classification.
- Interpretability: Shapelet-based features offer insights into the decision boundaries.
- Flexibility: Tunable parameters for different classification needs.
Wildboar provides a powerful alternative to gradient-based methods and is ideal for time series datasets where feature interpretability is essential.
To get started, select or upload a dataset and choose either Glacier or Wildboar as the classifier to generate counterfactual explanations for time series data.