Are Feature Selection and Feature Attribution the Same? A Comparative Survey -- Extended Version
Year: 2026
Type: preprint
Abstract: Feature Attribution methods in Explainable AI (XAI) such as Shapley Values (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) provide a Feature Importance value and thus resemble classic Feature Selection methods which calculate a score and form a ranking for features. Despite the fact that these methods belong to different fields of research, their technical similarities seem to be more than their discrepancies. Hence, we study in this paper the question: Are feature selection and feature attribution methods the same? We present a survey on the theoretical background and the design choices of feature selection and feature attribution methods. We provide a comprehensive unified evaluation framework to evaluate and compare feature selection and feature attribution methods using a combined set of metrics. We also provide studies and analyses in different settings to investigate different properties of feature selection and feature attribution methods. With this comparative survey, we aim to shed light on the different aspects of feature selection and feature attribution methods, their similarities, differences, and unique strengths and weaknesses, bridging the gap between the two groups of methods and bringing the fields together.
Source: Research Square
Authors: Muhammad Rajabinasab, Arthur Zimek
Institution: University of Southern Denmark
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204755770
