TimeREISE: Time Series Randomized Evolving Input Sample Explanation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Explainable AI

Background:

• Deep neural networks excel in classification but lack interpretability, limiting safety-critical applications.
• Existing explainable AI methods often are designed for imaging, not time series data.
• Interpretability is crucial for understanding and trusting AI decisions in critical domains.

Purpose of the Study:

• Introduce TimeREISE, a novel model-agnostic attribution method for time series classification.
• Evaluate TimeREISE's performance against existing interpretability techniques.
• Demonstrate TimeREISE's effectiveness in providing accurate and continuous explanations for time series models.

Main Methods:

• TimeREISE perturbs input data to generate attribution maps.
• The method analyzes attribution map characteristics like granularity and density.
• Evaluated using metrics such as deletion/insertion tests, Infidelity, Sensitivity, and explanation continuity.

Main Results:

• TimeREISE significantly outperforms existing methods on key interpretability benchmarks.
• Achieved superior performance in deletion/insertion tests, Infidelity, and Sensitivity.
• Demonstrated superior explanation continuity while preserving attribution map correctness.

Conclusions:

• TimeREISE offers a robust and versatile solution for time series classification interpretability.
• The method is model-agnostic, requires no prior data knowledge, and scales effectively.
• TimeREISE is suitable for diverse time series use cases regardless of data characteristics.