SCOPUS

[배혜림, 심성현] A Robust eXplainable Feature Selection Method Based on Order Statistics

영문초록: Recently, Feature Selection (FS) methods have been extensively researched
using eXplainable Artificial Intelligence (XAI). Among these methods, SHapley Additive
exPlanations (SHAP) is a representative approach. SHAP evaluates the impact of feature
subset on predicted values based on game theory. Consequently, the feature importance
can vary when iterated, and change depending on the prediction model used. In this
paper, we propose a robust FS method based on order statistics. We determine feature
ranking through two approaches: Feature Importance (FI) derived from the model-specific
using different prediction models and model-agnostic using iterative experiments. Finally,
we construct feature subsets based on the sum of these rankings and criteria. Through
experiments, we validate the robustness and efficiency of our approach. By utilizing this
approach, we can identify suitable feature subsets without the need to explore various FS
methodologies, regardless of the prediction model and data dimension.