Evaluation of interpretability methods for multivariate time series forecasting.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Time Series Analysis

Background:

• Interpreting machine learning model predictions is crucial, especially for local interpretability, which explains individual predictions.
• Existing research on interpretable AI has largely overlooked local interpretability methods for time series forecasting, focusing instead on classification tasks.

Purpose of the Study:

• To address the gap in local interpretability evaluation for time series forecasting.
• To propose and validate novel metrics for assessing the fidelity of local explanation methods in time series forecasting.

Main Methods:

• Introduction of two new evaluation metrics: Area Over the Perturbation Curve for Regression (aUPCR) and Ablation Percentage Threshold (APT).
• Theoretical extension and experimental validation of aUPCR and APT on four diverse time series datasets.
• Comprehensive comparison of various local explanation methods using the proposed metrics.

Main Results:

• The proposed metrics effectively measure the local fidelity of explanation methods in time series forecasting.
• Both aUPCR and APT facilitate a thorough comparison between different local explanation techniques.
• The metrics offer an intuitive framework for interpreting complex model predictions.

Conclusions:

• The developed metrics provide a robust and intuitive way to evaluate and compare local interpretability methods in time series forecasting.
• This work advances the field of interpretable AI by offering essential tools for time series-specific model interpretation.
• The findings support the broader application and development of interpretable machine learning in forecasting contexts.