SHAP을 활용한 음성 감정 인식의 설명 가능한 특징 선택

초록

Speech Emotion Recognition systems often face the curse of dimensionality when dealing with high-dimensional acoustic feature spaces. This paper presents an iterative feature selection approach using SHapley Additive exPlanations to identify the most discriminative features for emotion classification. We evaluate three tree-based classifiers - Decision Tree, Random Forest, and XGBoost on the CREMA-D dataset focusing on three primary emotions: anger, happiness, and sadness. Results demonstrate that integrating explainable AI techniques into iterative model training improves discrimination accuracy while providing interpretable insights into acoustic emotion indicators.