Toward More Explainable Nonlinear Dimensionality Reduction: A Feature-Driven Interaction Approach.

Area of Science:

• Data Science
• Computer Vision
• Machine Learning

Background:

• Nonlinear dimensionality reduction (NDR) is crucial for high-dimensional data visualization.
• Current NDR methods often lack explainability, hindering the interpretation of projected patterns in relation to original features.
• Existing interactive techniques inadequately integrate user input with the feature space, limiting insight generation.

Purpose of the Study:

• To develop a bidirectional interaction method that bridges the feature space and projections in NDR.
• To enable intuitive exploration of how feature weights influence data embeddings.
• To facilitate structured pattern discovery through automated, query-based interactions and visual semantics.

Main Methods:

• A bidirectional interaction approach allowing users to adjust feature weights directly.
• Definition of visual semantics to quantify projection changes.
• Utilization of a neural network to approximate the NDR projection process for enhanced scalability and responsiveness.
• Quantitative analysis of accuracy and scalability, complemented by a user study.

Main Results:

• The proposed method effectively links feature space manipulations to projection changes.
• A neural network approximation improved scalability and maintained accuracy in NDR.
• User studies demonstrated enhanced hypothesis generation and exploratory task performance with real-world data.

Conclusions:

• The bidirectional interaction method significantly improves the explainability of NDR techniques.
• The approach supports diverse analytical scenarios, enabling users to better explore and interpret high-dimensional data.
• Interactive exploration grounded in the feature space is key to unlocking deeper insights from complex datasets.