FADEx: Feature Attribution and Distortion-based Explanation of Dimensionality Reduction
Authors
Lucas Greff Meneses (ICMC-USP, University of Sao Paulo), Evandro S. Ortigossa (Weizmann Institute of Science), Claudio Silva (New York University), Luis Gustavo Nonato (University of Sao Paulo)
Presentation
- Session
- I don't trust you, explain yourself!
- Time
- Friday, Nov 13, 09:00 – 09:12 (US/Eastern) · session 08:00 – 09:30
- Location
- Hall Essex north
Links
Sign in to access the preprint PDF.
Sign in- Download Supplemental Material
Keywords
Dimensionality Reduction, Explainability, Local Feature Attribution
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 %black-box opaque models. However, non-linear DR techniques %operate as black boxes often function as opaque transformations themselves, making it challenging to understand how individual features influence instance positioning in the reduced space. This lack of transparency complicates the analysis and interpretation of structural patterns, hindering the ability to reason about the organization of high-dimensional data based on the projected layout. In order to address this challenge, dimensionality reduction explanation methods have shown promise in improving the understanding of the observed groups and cluster structures. Unfortunately, existing DR explanation approaches tend to suffer from limitations such as multiple attributions per feature and restricted applicability to specific dimensionality reduction methods, which hinder their use. In this work, we propose FADEx, a novel local per-instance feature attribution method that leverages local linear approximation via first-order Taylor expansion and Singular Value Decomposition to provide explanations. FADEx computes the local linear models via weighted least squares, eliminating the need for out-of-sample data mapping, making it agnostic to the DR method, while simultaneously providing local feature attributions and distortion analysis. Through qualitative and quantitative evaluations, comparisons with existing methods, and case studies, we demonstrate FADEx's effectiveness and versatility in providing explanations and analytical resources for analyzing the behavior of DR methods. The results indicate FADEx yields robust and reliable explanations, outperforming existing approaches in several aspects.
For Practitioners
FADEx is relevant to practitioners who rely on dimensionality reduction to explore and communicate high-dimensional data but need to understand why a projection looks the way it does, for example, data scientists and machine learning engineers inspecting learned embeddings, bioinformaticians and computational biologists interpreting cluster structure in single-cell or genomic data, financial and health analysts working with feature-rich tabular data, and visualization tool builders developing interactive systems on top of DR layouts. They can apply FADEx in three concrete ways: its feature attribution reveals which original features drive a point's position or a cluster's formation, turning an opaque scatter plot into an interpretable one; its SND distortion measure flags regions where the mapping has stretched or compressed neighborhoods, warning which parts of the layout are trustworthy and which could lead to misreadings of density or proximity; and its Feature Influence Vectors show how a point would move under a change in a given feature. Because FADEx is model-agnostic and works with any DR technique without retraining or out-of-sample mapping, practitioners can adopt it as a drop-in analytical layer on top of the DR methods they already use.