LEL: Lipschitz Continuity Constrained Ensemble Learning for Efficient EEG-Based Intra-subject Emotion Recognition

TL;DR

LEL addresses instability and noise sensitivity in EEG-based intra-subject emotion recognition by imposing Lipschitz continuity on spectral extraction, attention, and normalization modules. It introduces LGCBE, LGCN, and LGCA along with a learnable four-branch ensemble fusion to integrate temporal, spectral, spatial, and band-specific information. On EAV, FACED, and SEED, LEL achieves accuracies of $74.25\%$, $81.19\%$, and $86.79\%$, respectively, demonstrating superior robustness and generalization to intra-subject variability and noise. This Lipschitz-constrained ensemble approach enables more reliable real-time EEG emotion monitoring and provides a practical pathway for robust, subject-specific affective computing.

Abstract

Accurate and efficient recognition of emotional states is critical for human social functioning, and impairments in this ability are associated with significant psychosocial difficulties. While electroencephalography (EEG) offers a powerful tool for objective emotion detection, existing EEG-based Emotion Recognition (EER) methods suffer from three key limitations: (1) insufficient model stability, (2) limited accuracy in processing high-dimensional nonlinear EEG signals, and (3) poor robustness against intra-subject variability and signal noise. To address these challenges, we introduce Lipschitz continuity-constrained Ensemble Learning (LEL), a novel framework that enhances EEG-based emotion recognition by enforcing Lipschitz continuity constraints on Transformer-based attention mechanisms, spectral extraction, and normalization modules. This constraint ensures model stability, reduces sensitivity to signal variability and noise, and improves generalization capability. Additionally, LEL employs a learnable ensemble fusion strategy that optimally combines decisions from multiple heterogeneous classifiers to mitigate single-model bias and variance. Extensive experiments on three public benchmark datasets (EAV, FACED, and SEED) demonstrate superior performance, achieving average recognition accuracies of 74.25%, 81.19%, and 86.79%, respectively. The official implementation codes are available at https://github.com/NZWANG/LEL.

Figures (11)

• Figure 1: (a) LEL framework: Four lightweight branches fused for final emotion classification. (b) LGCBE: Lipschitz-constrained multi-band convolution and channel energy pooling. (c) LGCN: Lipschitz-constrained normalization. (d) LGCA: Lipschitz-constrained attention.
• Figure 2: Confusion Matrices of LEL (Our Model) Across Three Datasets: EAV (a), FACED (b), and SEED (c).
• Figure 3: Six scatter plots comparing emotion classification across combinations of Lipschitz constants (0.1, 1.0, 10.0) and training epochs (30, 50, 100), with raw EEG in (f). Color-coded by nine emotion categories, showing parameter effects on feature space separation and clustering.
• Figure 4: Partitioning with contextual information.
• Figure 5: Partitioning without contextual information.