Benchmarking Attention-Based Interpretability of Deep Learning in Multivariate Time Series Predictions.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Deep learning models in safety-critical systems require more than prediction accuracy; they need interpretable and robust explanations.
• Attention mechanisms are commonly used in deep neural networks for sequence modeling to provide intrinsic interpretability.

Purpose of the Study:

• To design diagnostic datasets for evaluating attention-based deep learning models in multivariate forecasting tasks.
• To assess the prediction performance, interpretability correctness, and sensitivity analysis of these models.

Main Methods:

• Development of a novel benchmark with synthetically generated datasets featuring time series interactions of increasing complexity.
• Empirical evaluation of attention-based deep neural networks using the benchmark across prediction performance, interpretability, and sensitivity.

Main Results:

• Most evaluated models demonstrated satisfactory prediction performance but often lacked correct interpretability.
• The IMV-LSTM model was the only one to achieve both high prediction performance and correct interpretability.
• IMV-LSTM effectively captures both autocorrelations and crosscorrelations in multivariate time series.

Conclusions:

• Attention mechanisms in deep learning models for forecasting do not inherently guarantee correct interpretability despite good prediction scores.
• IMV-LSTM shows promise for interpretable multivariate time series forecasting, with interpretability improving on more complex datasets.
• The proposed benchmark is crucial for diagnosing and improving the interpretability of attention-based deep learning models.