GACNet: A graph attention-based neural network for electroencephalography localization via contrastive learning

The Electroencephalography (EEG) based eye tracking methodology has encountered numerous obstacles due to the noisy, high-dimensional, and non-stationary characteristics inherent in brain data. In this study, we proposed a novel regression framework, named GACNet, which synergistically integrates a Graph Attention Network (GAT) for dynamic spatial modeling and supervised contrastive learning to enhance regression accuracy. For the encoder representation, we implement a temporal module that utilizes a 1D convolutional layer, channel-wise attention, and a bidirectional long short-term memory (Bi-LSTM) network. This configuration effectively captures both local and long-range temporal dependencies. Concurrently, the GAT-based spatial module treats the 129 EEG channels as a dynamic graph. It autonomously learns the importance of inter-channel relationships (edges) and node significance, effectively filtering irrelevant nodes and focusing on the most salient neural pathways for eye tracking. To further bolster the quality of representation, we employ a clustering-guided supervised contrastive (SupCon) loss, which ensures compactness among positive clusters while maximizing the separation from negative samples within the learned feature space. This is coupled with a joint optimization strategy that integrates regression loss, enabling the model to cultivate discriminative and task-relevant representations concurrently. Experimental results demonstrate that our proposed framework achieves root mean square errors (RMSE) of 17.21 mm, 20.18 mm, and 25.14 mm, substantially outperforming the previous state-of-the-art results of 24.37 mm, 27.91 mm, and 33.03 mm on the EEGEyeNet, EEGEyeTrack-Level 1, and EEGEyeTrack-Level 2 datasets, respectively. This remarkable improvement demonstrates the superiority of our dynamic graph-based attention model, combined with contrastive learning, over static or transformer-based approaches for robust EEG regression. The source code is publicly available at: https://github.com/timeseriessignal/GACNet.