Title: Navigating Neural Rhythms: A Morlet Spectral Transformer Approach to Cross-Subject EEG Emotion Decoding

Abstract: This research addresses the significant challenge of cross-subject emotion recognition from EEG data, a critical yet difficult objective within the field of brain-computer interfaces. In contrast to tasks characterized by distinct waveform signatures, emotional states in EEG signals are predominantly reflected in spectral power. These signals are inherently weak, noisy, and exhibit substantial variability between individuals. Current methodologies typically depend on either massive pretrained EEG foundation models, which demand extensive datasets and still face difficulties with inter-subject differences, or frequency-domain encoders that, while better aligned with spectral structures, are hindered by representation mismatches, tokenization dominated by drift, and an absence of spatial modeling specific to frequency bands. To overcome these limitations, we introduce the Morlet Spectral Transformer (MST), a novel architecture integrated with a spatiotemporal Transformer backbone and defined by three core innovations. First, we employ Morlet wavelet tokenization to generate a time-frequency representation that aligns with the multi-scale nature of brain rhythms, thereby adapting classical differential entropy features for Transformer compatibility. Second, we implement long-context baseline removal, a straightforward temporal normalization technique designed to eliminate subject-specific drift and redundancy in adjacent time windows. Third, we utilize frequency-specific spatial projection to learn distinct channel mixers for each frequency band, a strategy that captures interpretable, band-specific patterns and minimizes unwanted cross-channel interference. Our experiments demonstrate that MST surpasses both large-scale pretrained EEG foundation models and existing frequency-based methods across all SEED-family datasets, even without prior pretraining. These findings indicate that thoughtful design of data representations can provide a precise, economically viable, and transparent alternative to the resource-intensive approach of large-scale pretraining.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC