CoDrag: Coupled Label and Feature Prototype Dragging via Confidence-Aware Estimation for EEG-based Emotion Decoding.

Electroencephalogram (EEG) signals have been extensively investigated as a non-invasive modality with high temporal resolution for human emotion recognition. However, a fundamental challenge lies in the inherent mismatch between EEG features and emotional labels, a phenomenon we refer to as feature-label inconsistency. Such inconsistency primarily arises from the intrinsic non-stationarity of EEG data and their high susceptibility to noise. In addition, the instability of human emotional states, such as transition lags or insufficient immersion, further exacerbates the mismatch. To address this problem, we propose a discriminative regression framework termed coupled label and feature prototype dragging (CoDrag), which jointly refines soft-label assignment and latent feature structure through confidence-aware estimation. Specifically, CoDrag constructs label distributions via positive enhancement and negative suppression, promoting inter-class separability. Meanwhile, it adaptively learns emotion-specific prototypes guided by high-confidence samples while suppressing noisy or ambiguous ones. Extensive experiments on the SEED-IV, SEED-V, and MPED datasets demonstrate that CoDrag consistently outperforms existing methods in terms of classification accuracy, particularly under challenging cross-subject settings, while maintaining robustness against noisy labels and low-quality EEG data. Moreover, the proposed framework identifies a data-driven emotional prototype structure within the EEG feature space, exhibiting high consistency with the subjective emotion distance patterns defined by the dimensional modeling methods. Crucially, this finding provides objective evidence for structured emotional representations.