A Graph Attention Channel Aggregation Network via Multi-Level Representation Contrastive Learning for EEG-Based RSVP Tasks

Rapid serial visual presentation (RSVP)-based brain-computer interfaces (BCIs) offer a promising technique for visual target detection. However, existing methods for RSVP tasks suffer from reduced computational efficiency and detection accuracy due to difficulties in capturing discriminative EEG multi-scale spatial-temporal features. Furthermore, current methods often overlook the class balance between target and non-target samples, resulting in overfitting to multi-class samples. To address these limitations, we develop a novel graph attention channel aggregation network via multi-level representation contrastive learning (GACA-MLRCL) for EEGBased RSVP classification. Specifically, we first design a graph attention-based channel aggregation block that reduces noise and redundant information in spatial-temporal features by attention-weighted aggregation of multi-channel EEG spatialtemporal characteristics. Next, to reduce the computational complexity of traditional Transformers, we develop a spatialtemporal feature extraction block based on a global-local Transformer, which efficiently captures high-order discriminative spatial-temporal features. Additionally, we propose a multi-level representation contrastive learning strategy to overcome class imbalance. Competitive experimental results on a public dataset and a self-collected dataset demonstrate the effectiveness of our proposed GACA-MLRCL, indicating that our GACA-MLRCL achieves superior detection performance.