Class-Constrained t-SNE: Combining Data Features and Class Probabilities.

Area of Science:

• Machine Learning
• Data Visualization
• Computer Science

Background:

• Evaluating machine learning models often involves analyzing data features and class probabilities separately.
• Existing dimensionality reduction (DR) methods typically focus on only one of these perspectives.
• Integrating both data features and class probabilities in DR is challenging but crucial for comprehensive analysis.

Purpose of the Study:

• To develop a novel dimensionality reduction approach that combines data features and class probabilities into a unified visualization.
• To enable more effective model evaluation and interactive labeling by leveraging both data and probability information.
• To provide users with control over the balance between data features and class probabilities in the DR output.

Main Methods:

• Proposes class-constrained t-SNE, a new dimensionality reduction technique.
• Combines data features and class probabilities by optimizing a cost function with two components: data point positions and class landmarks.
• Introduces an interactive user-adjustable parameter to balance the influence of data features and class probabilities.

Main Results:

• Successfully integrates data features and class probabilities within a single DR result.
• Demonstrates application potential in model evaluation and visual-interactive labeling.
• Comparative analysis validates the effectiveness of the proposed DR approach.

Conclusions:

• Class-constrained t-SNE offers a unified perspective for analyzing data features and class probabilities.
• The method enhances model evaluation and facilitates interactive labeling through integrated visualization.
• User control over perspective weighting preserves mental maps and allows focused analysis.