Local Prediction Models for Spatiotemporal Volume Visualization.

Area of Science:

• Data Science
• Machine Learning
• Scientific Visualization

Background:

• Analyzing complex spatiotemporal data is challenging.
• Existing methods often require domain-specific knowledge or assumptions about data patterns.

Purpose of the Study:

• To develop a novel machine learning approach for detecting and visualizing complex behaviors in spatiotemporal volumes.
• To identify regions with unique, uncertain, or complex temporal dynamics.

Main Methods:

• Trained machine learning models to predict future data values based on historical neighborhood data.
• Evaluated prediction accuracy to identify areas of high error, indicating complex behavior.
• Utilized models of varying capacities to detect complexities at different scales.
• Aggregated and visualized prediction errors as time series or spatial volumes.

Main Results:

• Successfully detected and visualized spatiotemporal regions exhibiting complex, unique, or uncertain behaviors.
• Demonstrated the ability to highlight differences in unpredictable behavior across models of varying complexity.
• Showcased applications in adaptive timestep selection and ensemble dissimilarity analysis.
• Achieved meaningful results across diverse datasets with minimal data assumptions.

Conclusions:

• The proposed machine learning approach effectively identifies and visualizes complex spatiotemporal dynamics.
• The method is versatile, applicable to various domains with minimal prior assumptions.
• Prediction error serves as a robust indicator of interesting and complex data behaviors.