Feature importance-based interpretation of UMAP-visualized polymer space.

Area of Science:

• Data Science
• Machine Learning
• Bioinformatics

Background:

• Principal Component Analysis (PCA) is a linear dimensionality reduction (DR) technique, effective for linear data but limited for non-linear datasets.
• Interpreting complex, non-linearly reduced data remains a challenge in exploratory data analysis.

Purpose of the Study:

• To develop a novel method for enhancing the interpretability of data processed by non-linear DR techniques.
• To provide interpretable feature importance (FI) visualizations for complex datasets.

Main Methods:

• Applied density-based clustering to non-linear dimensionally reduced data.
• Utilized Random Forest (RF) classifiers to label clusters and calculated feature importance (FI).
• Employed Spearman's rank correlation, Gaussian process regression, and Boruta feature selection for enhanced interpretation and visualization.

Main Results:

• The proposed method generated interpretable FI-based images for handwritten digits and polymer datasets.
• Incorporating signed FI and using Gaussian process regression improved interpretation and visualization.
• Boruta feature selection effectively identified key features for cluster interpretation.

Conclusions:

• The developed technique successfully enhances the interpretability of non-linear DR results.
• The method is adaptable and effective for diverse datasets, including chemical and image data.
• Automation of the method yielded indicative results, demonstrating its practical applicability.