A Perception-Driven Approach to Supervised Dimensionality Reduction for Visualization.

Area of Science:

• Computer Science
• Data Visualization
• Machine Learning

Background:

• Dimensionality reduction (DR) is crucial for visualizing high-dimensional data.
• Existing methods like PCA and LDA have limitations with complex class structures and human perception.
• Maximizing class separability is key for effective data exploration.

Purpose of the Study:

• To develop a perception-driven linear DR approach that maximizes perceived class separation.
• To address the limitations of current DR methods in handling complex data structures.
• To create a DR technique that aligns with human perceptual capabilities.

Main Methods:

• Utilized recent perception-based separation measures, extended to be density-aware.
• Incorporated these measures into a customized simulated annealing algorithm for rapid projection generation.
• Evaluated the approach against state-of-the-art DR methods on 93 datasets.

Main Results:

• The perception-driven DR approach demonstrated superior performance in maximizing perceived class separation.
• Quantitative measures and human judgments confirmed the effectiveness of the new method.
• Successful case studies were conducted on class-imbalanced and unlabeled data.

Conclusions:

• The proposed perception-driven linear DR method effectively enhances class separation in data projections.
• This approach offers a significant improvement over traditional DR techniques for complex datasets.
• The method shows promise for broader applications, including imbalanced and unlabeled data analysis.