Hybrid Deep Learning Framework for Emotion Classification of BCI Data Using Autoencoder-Based Feature Extraction and Stacked Ensemble Learning

Being a vital component, Emotion classification is used in several applications such as mental health monitoring, human-computer interaction, and affective computing. Recently, machine learning and deep learning techniques have been successfully applied to emotion classification in Brain-Computer Interface (BCI) systems for attaining improved results as compared to traditional classification approaches. This paper introduces a novel hybrid Deep Learning framework that employs an Autoencoder-based feature extraction with a stacked Ensemble learning approach for enhancing classification performance. Herein, an autoencoder is utilized as an unsupervised feature extractor for learning a compact latent representation of EEG-based emotion data. Then, the extracted features are fed into two-base classifiers, namely Random Forest and Deterministic Weight Medication-Extreme Learning Machine. The resultant predictions are then combined using a meta-learning model (Logistic Regression). This means that the hierarchical learning strategy improves the robustness and generalization capabilities of the model. In this way, an accuracy of 98.126% is obtained using the proposed framework for emotion classification. This demonstrates the framework's effectiveness in capturing complex emotion patterns for EEG classification. The attained results showcase the significant improvement in classification performance achieved by integrating deep feature extraction and an ensemble-based hybrid model as compared with the standalone models.