EEG-Based Emotion Recognition Using Regularized Graph Neural Networks (1907.07835v4)

Abstract: Electroencephalography (EEG) measures the neuronal activities in different brain regions via electrodes. Many existing studies on EEG-based emotion recognition do not fully exploit the topology of EEG channels. In this paper, we propose a regularized graph neural network (RGNN) for EEG-based emotion recognition. RGNN considers the biological topology among different brain regions to capture both local and global relations among different EEG channels. Specifically, we model the inter-channel relations in EEG signals via an adjacency matrix in a graph neural network where the connection and sparseness of the adjacency matrix are inspired by neuroscience theories of human brain organization. In addition, we propose two regularizers, namely node-wise domain adversarial training (NodeDAT) and emotion-aware distribution learning (EmotionDL), to better handle cross-subject EEG variations and noisy labels, respectively. Extensive experiments on two public datasets, SEED and SEED-IV, demonstrate the superior performance of our model than state-of-the-art models in most experimental settings. Moreover, ablation studies show that the proposed adjacency matrix and two regularizers contribute consistent and significant gain to the performance of our RGNN model. Finally, investigations on the neuronal activities reveal important brain regions and inter-channel relations for EEG-based emotion recognition.

• The paper introduces a novel RGNN method that models inter-channel EEG relationships, addressing cross-subject variations and noisy labels.
• It employs an adjacency matrix inspired by brain connectivity alongside NodeDAT and EmotionDL regularizers to enhance classification accuracy.
• Experiments on SEED and SEED-IV datasets demonstrate robust performance, especially in beta and gamma frequency bands.

EEG-Based Emotion Recognition Using Regularized Graph Neural Networks

The paper "EEG-Based Emotion Recognition Using Regularized Graph Neural Networks" presents a novel approach for emotion recognition via Electroencephalography (EEG), leveraging the capabilities of Regularized Graph Neural Networks (RGNN). The proposition is centered on addressing some critical challenges in EEG-based emotion recognition, including inadequate modeling of EEG channel topology, significant cross-subject EEG variations, and the presence of noisy labels.

The researchers introduce RGNN as a method that utilizes the inherent topology of EEG channels by employing a graph neural network framework. This approach models the inter-channel relations via an adjacency matrix inspired by the biological organization of the human brain. The adjacency matrix aids in capturing not only local but also global relations among EEG channels, supported by the incorporation of both anatomical connectivity and functional connectivity principles.

Furthermore, RGNN introduces two specific regularizers to better handle cross-subject variations and label noise. Node-wise Domain Adversarial Training (NodeDAT) is applied to reduce domain discrepancies across subjects, thus enhancing subject-independent classification accuracy. Emotion-aware Distribution Learning (EmotionDL) serves to manage noisy labels, which frequently arise due to the variability in human emotional responses to stimuli.

Complementing the theoretical propositions, the paper methodically evaluates RGNN against existing state-of-the-art models using two comprehensive EEG datasets: SEED and SEED-IV. The results demonstrate that RGNN consistently outperforms other models in both subject-dependent and subject-independent scenarios, achieving significant gains especially in the latter, thus affirming its enhanced cross-subject robustness.

An in-depth investigation into the performance across different EEG frequency bands reveals that RGNN performs optimally on beta and gamma bands, aligning with previous findings in EEG literature regarding their relevance in emotional processing. The ablation studies further corroborate the contribution of the proposed adjacency matrix and regularizers toward the enhanced performance of RGNN.

The implications of this research span both practical and theoretical domains. Practically, RGNN presents a sophisticated methodological advancement in the field of affective computing, potentially enriching the robustness and accuracy of emotion-recognition applications across various domains, including personalized medicine and human-computer interaction. Theoretically, this work contributes a structured model that elegantly encapsulates the intricate topology of EEG data within the framework of graph neural networks, paving the way for further explorations in brain-inspired AI model development.

Future investigations may seek to refine the domain adaptation techniques used in RGNN and explore its application to fewer-channel EEG data, thereby expanding its utility and applicability. Additionally, the integration of spatial filtering techniques to enhance EEG spatial resolution represents another promising avenue for exploration. By addressing these areas, RGNN can further consolidate its position as a frontrunner in EEG-based emotion recognition research.