INSIGHTS: New Method for Global Time-Series Model Explanations
A new approach called INSIGHTS has been developed by researchers, designed to provide global explanations for time series models in a user-focused, model-agnostic manner. Unlike traditional techniques that emphasize local attributions for specific instances, INSIGHTS creates sample summaries that deliver a holistic view of model performance. This method employs utility functions to balance the significance and variety of time series samples, taking into account domain-specific factors like surpassing established norms. The emphasis is on clarity, efficiency, and transparency. Evaluation of INSIGHTS was conducted through experiments, user interviews, and a study, demonstrating its ability to generate diverse and informative subsets. Detailed findings are available in arXiv preprint 2605.18849.
Key facts
- INSIGHTS is a model-agnostic, user-centric approach for global explanations of time series models.
- It generates sample summaries to provide a comprehensive overview of model behavior.
- The method balances importance and diversity of time series samples.
- Utility functions capture domain-specific aspects like exceeding domain norms.
- Evaluation included experiments, interviews, and a user study.
- Results indicate INSIGHTS effectively constructs comprehensive, diverse subsets.
- The approach prioritizes simplicity, efficiency, and transparency.
- Current methods focus on local, instance-level attributions.
Entities
Institutions
- arXiv