Introduction
Dimensionality Reduction (DR) is a fundamental technique in visual cluster analysis, enabling the projection of high-dimensional data into interpretable 2D scatterplots. However, established methods like t-SNE and UMAP often suffer from visual overlap between clusters and provide limited user control. This paper introduces Contrastive Dimensionality Reduction (CDR), a parametric approach that leverages contrastive learning and a novel gradient redefinition technique to produce embeddings with superior cluster separation and support for human-in-the-loop interaction.
Methodology
CDR addresses the limitations of traditional DR through three primary innovations:
1. Contrastive Learning Framework
Unlike non-parametric methods that focus solely on distance preservation, CDR utilizes a neural network trained with a contrastive loss (NT-Xent). The model treats high-dimensional neighbors as positive pairs to be pulled together and non-neighbors as negative pairs to be pushed apart in the 2D embedding space.
2. Gradient Redefinition
A key contribution of this work is the "gradient redefinition" technique. The authors observed that standard contrastive loss often fails to create sufficient "white space" between clusters. By redefining the gradient of the loss function using a skewed distribution, CDR aggressively targets and separates points likely to belong to different clusters, even when they overlap in the initial projection.
3. Interactive Steering
As a parametric model, CDR can be fine-tuned in real-time. Users can provide domain knowledge by specifying "must-link" or "cannot-link" constraints between data points. The underlying neural network updates its weights based on these interactions, allowing the visualization to evolve according to human insights.
Evaluation and Results
The performance of CDR was evaluated against 11 other DR techniques across 10 diverse datasets, including image sets (e.g., Cifar10, MNIST) and tabular data (e.g., Indian Food, WiFi signals).
Quantitative Performance
CDR significantly outperformed t-SNE and UMAP on key metrics such as the Silhouette Coefficient (SC) and Distance Consistency (DSC). These results indicate that CDR produces tighter clusters with wider inter-cluster gaps, facilitating easier visual identification. As shown in the comparison above, CDR achieves the highest scores across Trustworthiness, Continuity, and Neighborhood Hit (NH) metrics.
User Study
A study involving 20 participants demonstrated that CDR improved cluster identification precision to 80.68%, compared to approximately 66% for t-SNE and UMAP. Participants reported higher confidence and lower cognitive load when using CDR-generated plots, particularly when utilizing the interactive steering features to resolve ambiguities.
Conclusion
Contrastive Dimensionality Reduction represents a significant advancement in visual data exploration. By prioritizing visual separation and enabling interactive steering, CDR bridges the gap between automated structure preservation and human-centric analysis. While the initial training phase is more intensive than some non-parametric methods, the resulting clarity and flexibility make it a powerful tool for complex data exploration tasks.
Summary updated