Visual Analysis of Class Separations With Locally Linear Segments.

Area of Science:

• Data Visualization
• Machine Learning Analysis
• High-Dimensional Data Exploration

Background:

• High-dimensional labeled data is prevalent in classification and clustering.
• Analyzing class separations and boundaries is crucial for understanding data.
• Traditional dimension reduction methods struggle with complex non-linear decision boundaries, leading to distortion and poor interpretability.

Purpose of the Study:

• To develop a visual analysis approach for exploring non-linear class separations in high-dimensional data.
• To enhance the interpretability and separability of complex decision boundaries.
• To leverage explainability from linear projections for non-linear structures.

Main Methods:

• Extracting locally linear segments to approximate non-linear separations.
• Utilizing multiple local projection results for exploration, unlike single scatterplots.
• Applying a visual analysis approach combining linear explainability with non-linear structures.

Main Results:

• Demonstrated effectiveness in case studies on two labeled datasets.
• Successfully visualized complex class separations that are challenging for traditional methods.
• Provided improved insights into non-linear decision boundary structures.

Conclusions:

• The proposed visual analysis approach effectively addresses limitations of traditional dimension reduction techniques.
• It enhances the exploration of complex, non-linear class separations through locally linear approximations.
• The method offers improved separability and interpretability for high-dimensional data analysis.