SIFIAE: An adaptive emotion recognition model with EEG feature-label inconsistency consideration

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• Affective overlap, where past emotions influence current states, is a challenge in electroencephalogram (EEG)-based emotion recognition.
• Short rest intervals in experiments cause neural responses to persist, leading to feature-label inconsistency in EEG data.
• This inconsistency, termed affective overlap, hinders accurate emotion recognition from EEG signals.

Purpose of the Study:

• To address the under-researched problem of affective overlap in EEG emotion recognition.
• To develop a novel model that can adaptively explore and mitigate the impact of sample inconsistency.
• To improve the performance of emotion recognition systems by considering both sample inconsistency and feature importance.

Main Methods:

• Introduction of a variable to adaptively explore sample inconsistency in EEG data.
• Proposal of a semi-supervised model named SIFIAE (Semi-supervised learning for joint sample Inconsistency and Feature importance exploration for Affective Emotion recognition).
• Development of an efficient optimization method tailored for the SIFIAE model.

Main Results:

• The SIFIAE model demonstrated effectiveness on the SEED-V dataset.
• Achieved average accuracies ranging from 67.01% to 73.26% across six cross-session emotion recognition tasks.
• Observed sample weights increasing at the start of trials, supporting the affective overlap hypothesis.

Conclusions:

• The SIFIAE model successfully accounts for feature-label inconsistency caused by affective overlap.
• Identified critical EEG bands and channels through feature importance analysis, outperforming models that ignore inconsistency.
• The findings highlight the importance of addressing affective overlap for robust EEG-based emotion recognition.