Attribute-based Explanation of Non-Linear Embeddings of High-Dimensional Data

Area of Science:

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

Background:

• High-dimensional data embeddings are crucial for data exploration and analysis.
• Explaining non-linear projections and their relation to input attributes is a significant challenge.
• Existing augmentation techniques for embedding visualization have limitations.

Purpose of the Study:

• To review and discuss limitations of current data augmentation techniques for embeddings.
• To introduce the Non-Linear Embeddings Surveyor (NoLiES) for enhanced visualization and analysis of non-linear projections.
• To demonstrate the utility of NoLiES in various complex data scenarios.

Main Methods:

• Development of rangesets, a novel set-based visualization strategy for binned attribute values.
• Integration of rangesets with interactive analysis in a small multiples setting.
• Exploration of the link between algebraic topology and the rangeset visualization approach.

Main Results:

• Rangesets enable users to quickly observe structure and detect outliers in projected data.
• NoLiES effectively addresses challenges including complex attribute distributions, numerous attributes, and large datasets.
• Successful application to a real-world thermodynamics problem for understanding latent features.

Conclusions:

• NoLiES provides an effective solution for interpreting non-linear data embeddings.
• Rangesets offer a powerful new method for visualizing and analyzing projected data structures.
• The tool facilitates deeper insights into complex datasets and machine learning models.