FADEx Explains Dimensionality Reduction with Feature Attribution

Lucas Greff Meneses, Evandro S. Ortigossa, Claudio Silva, Luis Gustavo Nonato· July 31, 2026 View original

Key takeaways

  • FADEx explains how features influence dimensionality reduction.
  • It provides local, per-instance feature attributions and distortion analysis.
  • The method is agnostic to the specific DR technique used.
  • FADEx enhances transparency and interpretability of complex data.

Who benefits

HealthcareFinancial ServicesMarketingResearch & DevelopmentCybersecurity

Summary

FADEx is a new method that explains how individual features influence data points in reduced-dimensional spaces, addressing the opacity of non-linear dimensionality reduction techniques. It provides local feature attributions and distortion analysis, applicable to various DR methods.

This paper introduces FADEx, a novel method for explaining dimensionality reduction (DR) techniques, particularly non-linear ones that often act as opaque transformations. DR is crucial for high-dimensional data exploration and model explanation, but understanding how specific features contribute to the positioning of instances in the reduced space has been challenging. FADEx tackles this by providing local, per-instance feature attributions and distortion analysis. The method utilizes local linear approximation via first-order Taylor expansion and Singular Value Decomposition, computing local linear models through weighted least squares. This approach makes FADEx agnostic to the specific DR method used and eliminates the need for out-of-sample data mapping. Qualitative and quantitative evaluations, including comparisons with existing methods, demonstrate FADEx's effectiveness and versatility in offering robust and reliable explanations for DR behavior.

Why it matters

Professionals working with complex machine learning models and high-dimensional data need transparent ways to understand how data is transformed, enabling better interpretation of patterns and more trustworthy AI systems.

How to implement this in your domain

  1. 1Integrate FADEx into existing data analysis and machine learning pipelines for improved interpretability of DR outputs.
  2. 2Apply FADEx to complex datasets to gain deeper insights into feature contributions in reduced spaces.
  3. 3Train data scientists and ML engineers on using FADEx to explain model latent spaces and data structures.
  4. 4Develop visualizations based on FADEx outputs to communicate DR explanations more effectively to stakeholders.
  5. 5Benchmark FADEx against current interpretability tools to assess its advantages for specific use cases.

Original post by Lucas Greff Meneses, Evandro S. Ortigossa, Claudio Silva, Luis Gustavo Nonato

"arXiv:2607.27463v1 Announce Type: new Abstract: Dimensionality Reduction (DR) is a fundamental tool for high-dimensional data exploration, reducing the complexity of latent spaces of machine learning models, and assisting in the explanation of complex opaque models. However, non-…"

View on X

Originally posted by Lucas Greff Meneses, Evandro S. Ortigossa, Claudio Silva, Luis Gustavo Nonato on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses