TimeTuner: Diagnosing Time Representations for Time-Series Forecasting with Counterfactual Explanations

Area of Science:

• Data Science
• Machine Learning
• Time-Series Analysis

Background:

• Deep learning (DL) models are widely used for time-series forecasting.
• Model success often relies on effective data representations and feature engineering.
• Automated feature learning methods struggle with prior knowledge integration and interaction identification.

Purpose of the Study:

• Introduce TimeTuner, a visual analytics framework for understanding DL time-series forecasting.
• Connect model behavior to time-series representations, correlations, stationarity, and granularity.
• Improve reliability and interpretability of DL models in forecasting tasks.

Main Methods:

• Utilize counterfactual explanations to link time-series representations, features, and predictions.
• Employ coordinated views: partition-based correlation matrix and bivariate stripes.
• Incorporate user interactions for transformation selection, feature space navigation, and performance reasoning.

Main Results:

• Demonstrated TimeTuner with smoothing and sampling transformations on sunspot and air pollutant data.
• Showcased the framework's ability to characterize time-series representations.
• Provided evidence that TimeTuner guides effective feature engineering processes.

Conclusions:

• TimeTuner offers a novel approach to visual analytics for deep learning time-series forecasting.
• The framework aids analysts in understanding the impact of data representations on model performance.
• TimeTuner facilitates more reliable and interpretable forecasting through guided feature engineering.