Attention-Based Temporal Graph Representation Learning for EEG-Based Emotion Recognition.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Electroencephalogram (EEG)-based emotion recognition leverages the objective nature of emotional expression in the central nervous system.
• Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have advanced EEG signal analysis but have limitations in capturing spatial information and temporal dependencies.

Purpose of the Study:

• To propose an advanced network, the attention-based temporal graph representation network (ATGRNet), for improved EEG-based emotion recognition.
• To overcome the limitations of existing CNN and RNN models in EEG analysis.

Main Methods:

• Implemented a hierarchical attention mechanism to integrate prioritized frequency band and channel features from EEG signals.
• Utilized a graph convolutional neural network with top-k operation to model inter-electrode relationships under various emotional states.
• Employed a residual-based graph readout mechanism for aggregating node-level EEG features into graph-level representations.
• Applied a temporal convolutional network (TCN) to extract temporal dependencies from the graph-level EEG features.

Main Results:

• The proposed ATGRNet demonstrated superior performance in EEG-based emotion recognition.
• Experimental results on SEED, DEAP, and FACED datasets confirmed ATGRNet's effectiveness.
• ATGRNet surpassed state-of-the-art graph-based methods for emotion recognition from EEG data.

Conclusions:

• ATGRNet offers a robust framework for EEG-based emotion recognition by effectively integrating spatial and temporal information.
• The novel network architecture addresses limitations of previous methods, paving the way for more accurate emotion detection.