소수 채널 EEG 감정 인식을 위한 클래스 조건부 확산 모델 기반 데이터 증강 기법

초록

This study proposes a diffusion-based data augmentation framework optimized for few-channel EEG emotion recognition, addressing the intrinsic data scarcity and information loss inherent in low-cost, portable EEG systems. Using the SEED dataset, we utilized 30-dimensional differential entropy (DE) features extracted from six frontal channels. To effectively model these low-dimensional tabular features, we designed a class-conditional ϵ-prediction Denoising Diffusion Probabilistic Model (DDPM) incorporating a FiLM-ResMLP architecture. Furthermore, we introduced a novel Mean-Guided Guidance Sampling mechanism to explicitly steer generated samples toward class-consistent statistical distributions, thereby enhancing training stability. Experimental results demonstrate that the proposed FiLMResMLP based Conditional DDPM achieves a classification accuracy of 57.87%, representing a significant improvement of 4.85%p over the baseline (53.02%). The proposed method consistently outperforms state-of-the-art generative baselines, including Conditional DDPM (FiLM-UNet), Conditional WGAN and VAED2GAN. Notably, our analysis reveals that while the conventional UNet architecture tends to overfit in the low-dimensional feature space, the proposed ResMLP structure exhibits superior robustness and statistical compatibility with tabular EEG data. By demonstrating that data augmentation can effectively compensate for information loss caused by channel reduction, this work presents a viable methodology for overcoming the performance limitations of practical few-channel BCI systems.

키워드