EEG-Based Emotion Recognition via Prototype-Guided Disambiguation and Noise Augmentation in Partial-Label Learning

Yiming Wang, Bin Zhang, Lamei Di

Keywords: EEG-based emotion recognition, partial label learning, self-distillation, Prototypes, noise augmentation

TL;DR: We introduce a semantic-based candidate label generation method, and a partial label learning model for EEG-based emotion recognition integrating prototype-guided self-distillation and controllable noise augmentation.

Abstract: EEG-based emotion recognition offers an objective method for diagnosing emotion-related health issues, but the inherent complexity of emotions often leads to annotation errors and noisy labels. To simulate this labeling process in emotion recognition, we propose a semantic-based candidate label generation method leveraging the GloVe vectors, which considers the semantic relationships between emotions. Under the Partial Label Learning (PLL) scenario, we introduce a novel model called PGNA-PL (Prototype-Guided Noise-Augmented Partial Label Learning). This model learns inter-class relationships of emotions using prototypes, and uses a self-distillation mechanism to iteratively guide the classifier's disambiguation process. To address the low signal-to-noise ratio (SNR) of EEG, we introduce a noise augmentation strategy inspired by the mixup method, incorporating controllable noise to enhance model robustness. Experiments on three public datasets (SEED, SEED-IV, SEED-V) show that our approach achieves state-of-the-art performance, surpassing existing PLL baselines across different candidate label generation modes. Our method effectively disambiguates complex emotions and shows promising results in assisting in the recognition of fear-related disorders.

Primary Area: applications to neuroscience & cognitive science

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10677