Application of Time Series Model LT-MAE in EEG Emotion Recognition

Abstract

Electroencephalography (EEG) signals are non-linear and non-stationary. Traditionally, they are segmented into time windows for feature extraction under the assumption of independence and identical distribution, ignoring temporal connections and distribution discrepancies. Additionally, generating high-quality annotations for dynamic emotions is labor-intensive and time-consuming. To address these issues, we propose LT-MAE, a self-supervised learning model. It segments EEG signals into continuous time steps and uses long short-term memory network (LSTM) to learn context representations across channels. Emotion distributions are learned using an enhanced mask autoencoder with classification and reconstruction tasks. This approach assesses emotional changes over continuous time steps to determine long-term emotional inclinations. Experiments on SEED-IV and DEAP datasets show that LT-MAE learns a broader time-step emotion distribution in the coding space, improving emotion detection accuracy and mitigating labeling inaccuracies due to finite-time granularity. Unlike traditional methods that assume independent and identically distributed data, LT-MAE captures temporal connections and distribution discrepancies. By leveraging LSTM and enhanced mask autoencoder techniques, it provides more accurate emotion recognition while reducing reliance on costly annotations. In conclusion, LT-MAE offers an effective solution to limitations in conventional EEG feature extraction. Using self-supervised learning, it enhances the understanding of temporal sentiments and improves emotion recognition accuracy, addressing challenges related to annotation quality and granularity.