Shape-Driven Coordinate Ordering for Star Glyph Sets via Reinforcement Learning.

Area of Science:

• Computational Statistics
• Machine Learning
• Data Visualization

Background:

• The coordinate ordering problem in star glyphs impacts the perception of class separability.
• Existing methods may not efficiently optimize glyph arrangement for visual analysis.

Purpose of the Study:

• To develop a neural optimization model for solving the coordinate ordering problem in star glyphs.
• To enhance the perception of class separability in multi-labeled datasets using optimized glyph arrangements.

Main Methods:

• Utilized shape context descriptors to measure perceptual distance between glyphs.
• Employed a silhouette coefficient to quantify class separability.
• Trained a neural network with reinforcement learning, incorporating an RNN encoder-decoder with attention, to find optimal coordinate orders based on silhouette coefficients.

Main Results:

• The proposed model successfully learned optimal coordinate orders that improve perceived class separation, as validated by user studies.
• Introduced a neural network for efficient shape context descriptor similarity estimation, accelerating training.
• Demonstrated robustness and generalization across various data sizes, dimensions, and class numbers, including unseen configurations.

Conclusions:

• The neural optimization model effectively addresses the star glyph coordinate ordering problem, improving visual class separation.
• The method is adaptable to other visualization types like RadViz by substituting the quality metric.
• This approach offers a scalable and effective solution for enhancing data visualization interpretability.