Genetic Programming for Evolving a Front of Interpretable Models for Data Visualization.

Area of Science:

• Data Mining
• Machine Learning
• Scientific Visualization

Background:

• Effective data visualization is crucial for understanding large datasets.
• Current advanced methods like t-distributed stochastic neighbor embedding (t-SNE) lack interpretability.
• Many scientific domains require visualizations linked to original data features.

Purpose of the Study:

• To develop a novel data visualization method that combines high visual quality with model interpretability.
• To address the opacity limitation of existing powerful visualization techniques.
• To enable deeper data insights by connecting visualizations to original features.

Main Methods:

• Proposing a genetic programming (GP) approach, termed GP-tSNE.
• Evolving interpretable mappings from datasets to visualizations.
• Employing a multiobjective optimization strategy to balance visual quality and model complexity.

Main Results:

• GP-tSNE successfully generates high-quality, interpretable visualizations.
• Comparative testing against baseline methods demonstrates superior data insight potential.
• The multiobjective approach yields a variety of visualizations offering different quality-complexity tradeoffs.

Conclusions:

• GP-tSNE offers a powerful solution for interpretable big data visualization.
• The method enhances data understanding by linking visualizations to original features.
• Multiobjective analysis provides richer insights through joint examination of multiple models.