Cross-Level Topological Framework: Learning Explainable Region-Channel Representations from EEG Signals for Emotional Decoding.

Constructing functional connectivity networks from electroencephalogram (EEG) channels and using graph neural networks for emotion recognition have emerged as a significant technical route in EEG emotion recognition. However, most existing approaches are limited to estimating brain graph networks based on EEG full-channel signals, failing to adequately explore the representations between and within brain regions. To address this limitation and further investigate the interactions between channels and regions, an explainable cross-level topological network (ECTN) is proposed for EEG emotion recognition, which is designed to capture EEG functional interactions from channel-level to region-level. Within the ECTN framework, three modules are designed, namely cross-region topological feature fusion module, specific-region position-guided attention module, and bidirectional gated fusion module. Specifically, EEG functional interactions are explicitly decoupled into two complementary views: global region interactions and local region dynamics. Additionally, the bidirectional gated fusion module leverages the inclusion relationships between channels and brain regions to further integrate region-level and channel-level features. The ECTN model is evaluated on the publicly available SEED series of datasets, SEED, SEED-IV and SEED-V. Experimental results indicate that our method achieves superior performance, effectively validating the benefits of exploring channel-wise and region-wise interactions.