Interactive Visual Cluster Analysis by Contrastive Dimensionality Reduction.

Area of Science:

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

Background:

• Dimensionality reduction is crucial for visual cluster analysis but faces challenges in maintaining neighborhood structures and enabling user interaction.
• Existing methods struggle to achieve clear visual cluster separation while preserving data topology.

Purpose of the Study:

• To introduce a novel contrastive dimensionality reduction (CDR) approach for enhanced interactive visual cluster analysis.
• To improve the quality of embeddings for better cluster separation and user-guided analysis.

Main Methods:

• Incorporated contrastive learning into dimensionality reduction to create high-quality embeddings.
• Redefined the loss function's gradient for negative pairs to boost visual cluster separation.
• Implemented link-based interactions for user steering of embeddings within a visual interface.

Main Results:

• CDR demonstrated superior performance in preserving neighborhood structures and enhancing visual cluster separation compared to existing techniques.
• Ablation experiments confirmed the effectiveness of the gradient redefinition strategy.
• User studies showed CDR outperformed t-SNE and UMAP in cluster identification tasks.

Conclusions:

• Contrastive dimensionality reduction (CDR) offers a powerful approach for interactive visual cluster analysis.
• CDR effectively addresses limitations of traditional methods by improving embedding quality and enabling user interaction.
• The proposed method facilitates more accurate and intuitive cluster identification from high-dimensional data.