CrCo- Mlgcn: A Cross-scale Co-learning Based Multi-Level Graph Convolutional Network for Brain-Computer Interface

Functional near-infrared spectroscopy (fNIRS) decoding is a crucial foundation for Brain-Computer Interface (BCI) technology. However, existing methods commonly concentrate on time-frequency features and overlook positional information, thereby failing to fully utilize the high spatial resolution advantages of fNIRS. Furthermore, some methods employ multi-branch structures, but they fail to consider the cooperative interaction between the branches, leading to suboptimal performance. To address these limitations, we propose a cross-scale co-learning based multilevel graph convolutional network (CrCo-MLGCN). Specifically, a multi-level graph convolution module is crafted with three branches to extract features from different levels, including local, regional, and global levels. Then, a cross-scale co-learning module is employed to harmonize the model's multiple branches, ensuring efficient parameter utilization and preventing redundancy. Experimental results on two public datasets demonstrate that the proposed CrCo-MLGCN outperforms current state-of-the-art approaches, confirming the effectiveness of our proposed method.