Interactive polar diagrams for model comparison.

Area of Science:

• Computational Science
• Data Science
• Scientific Modeling

Background:

• Conventional methods for evaluating complex models across diverse fields like biology, climatology, and machine learning are often inadequate when using multiple evaluation metrics simultaneously.
• Traditional tabular presentations of multi-model evaluation scores hinder the determination of model similarities and performance rankings.

Purpose of the Study:

• To introduce a novel approach for assessing and comparing the performance of complex models.
• To provide an accessible tool for visualizing model performance and relationships.

Main Methods:

• The study combines statistics, information theory, and data visualization techniques.
• Juxtaposed Taylor and Mutual Information Diagrams are employed to track and summarize model performance.
• The approach allows for the visualization of linear and nonlinear relationships between models.

Main Results:

• The developed library, polar-diagrams, offers the first publicly available interactive implementations of Taylor Diagrams and Mutual Information Diagrams.
• These diagrams are extended to display temporality, multimodality, multivariate datasets, and scalar model properties like uncertainty.
• The library supports both continuous and categorical data attributes.

Conclusions:

• The polar-diagrams library provides a user-friendly solution for rapidly assessing the performance of complex models.
• It is applicable across various domains including machine learning, climate science, and biomedical research.