Robust Latent Multi-Source Adaptation for Encephalogram-Based Emotion Recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• Subject-independent electroencephalogram (EEG) classifiers face performance degradation due to diverse subject-specific EEG patterns.
• Cross-dataset and cross-subject emotion recognition remains a challenge in EEG-based machine learning.

Purpose of the Study:

• To develop a robust framework for cross-subject and cross-dataset emotion recognition using EEG signals.
• To address the performance degradation of subject-independent classifiers by uncovering domain-invariant latent subspaces.

Main Methods:

• Proposed a Latent Multi-source Adaptation (LMA) framework.
• Uncovered multiple domain-invariant latent subspaces by aligning statistical and semantic distribution discrepancies.
• Employed a novel low-rank regularization term to leverage correlated knowledge among multiple data sources.

Main Results:

• Demonstrated superior or comparable performance of the LMA framework against state-of-the-art methods.
• Validated the framework's effectiveness on DEAP and SEED datasets for EEG-based emotion recognition.
• Showcased the ability to train multiple domain-invariant classifiers collaboratively within a unified framework.

Conclusions:

• The LMA framework offers a robust solution for cross-subject and cross-dataset EEG emotion recognition.
• The proposed method effectively mitigates performance degradation caused by subject variability.
• LMA successfully utilizes correlated knowledge across multiple data sources for improved emotion recognition accuracy.