Robust explainer recommendation for time series classification

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Time series classification is vital in fields like human activity recognition and sports analytics.
• Explainability is increasingly important for understanding time series data and classification models.
• Existing explanation techniques produce saliency maps that can conflict, creating uncertainty about their reliability.

Purpose of the Study:

• To introduce a novel framework for quantitatively evaluating and ranking explanation methods in time series classification.
• To enable robust comparison of different explanation techniques and recommend the most suitable one for a given dataset.
• To address the challenge of conflicting saliency maps from various explanation methods.

Main Methods:

• Proposing AMEE (Model-Agnostic Explanation Evaluation), a framework for recommending saliency-based explanations.
• Utilizing data perturbation guided by explanations to assess their informativeness and impact on classification accuracy.
• Aggregating accuracy loss across diverse perturbations and classifiers for robust evaluation.

Main Results:

• Perturbing discriminative time series segments identified by explanations significantly alters classification accuracy.
• The AMEE framework successfully ranks explanation methods, outperforming random and oracle baselines.
• Demonstrated effectiveness across synthetic, diverse time-series datasets, and a real-world case study.

Conclusions:

• The proposed AMEE framework provides a robust method for evaluating and selecting explanation techniques for time series classification.
• Quantitative evaluation of explanation methods is feasible and essential for reliable model interpretability.
• This work facilitates the selection of optimal explainers, enhancing trust and understanding in time series classification models.