TASTE: Trait-like and State-like Two-stream Framework for EEG Imputation

Electroencephalography (EEG) is a core sensing modality for brain-computer interfaces, yet real-world recordings often contain missing values due to motion artifacts, electrode issues, and device or channel failures. Unhandled missingness can bias analysis and degrade downstream decoding. Existing deep imputation methods typically formulate EEG recovery as a generic spatiotemporal completion problem and thus underutilize EEG-specific priors. Motivated by neurophysiological evidence, we characterize EEG signals with complementary trait-like regularities (reproducible temporal/spatial patterns) and state-like variations (acquisition noise and informative nonstationary dynamics). Based on this view, we propose TASTE, a Trait-like And State-like Two-stream framework for EEG imputation. TASTE integrates (i) a Multi-View Trait Modeling module that learns persistent temporal, spatial, and spatio-temporal priors via a tri-branch learnable codebook with squeeze-and-excitation fusion, and (ii) a State-Aware Reconstruction module that performs noise-robust feature-space completion and preserves meaningful nonstationarity using de-stationary attention. We evaluate TASTE on five public EEG benchmarks spanning diverse tasks. TASTE consistently improves imputation fidelity and downstream classification, achieving the best performance in most settings and delivering average MAE/MSE gains of 22.22%/28.60% over strong baselines. Source code is available at https://github.com/XJTU-EEG/TASTE.