- The paper introduces a two-stage SSAS framework that dynamically selects beneficial source domains to mitigate negative transfer.
- The methodology integrates adversarial loss, MMD, and Gaussian noise injection to enhance model robustness, achieving up to 91.97% accuracy.
- Ablation studies and t-SNE visualizations validate that SSAS improves class clustering and reduces inter-subject variability in EEG signals.
Cross-Subject EEG-Based Emotion Recognition with Source Selection and Adversarial Strategy
Electroencephalography (EEG)-based affective brain-computer interfaces (aBCIs) are central to emotion recognition, yet generalization across subjects is persistently challenged by inter-individual variability and negative transfer. Inter-subject differences in neuroanatomy, cognitive state, and signal acquisition compound the difficulty of adapting models learned from one set of subjects to another. Classical cross-subject transfer learning techniques often overlook the selection bias inherent in multi-source transfer, leading to suboptimal performance, especially when negative transfer from dissimilar subjects dominates. Traditional approaches either align distributions at feature level (MMD-based), incorporate graph neural architectures to resolve spatial or temporal structure, or use adversarial learning (DANN and derivatives). None directly resolve the issue that some source domains are fundamentally detrimental for adaptation—an observation supported by empirical failure of similarity metrics (e.g., MMD) to predict transfer efficacy.
The SSAS Framework: Adversarial Source Selection and Adaptation
SSAS (Source Selection with Adversarial Strategy) introduces an explicit two-stage pipeline: Source Selection (SS) and Adversarial Strategies (AS). The SS module leverages domain labels not for alignment but to amplify separability, adversarially inflating domain differences for the primary purpose of quantifying source domain transferability. By intentionally disrupting class separability while maximizing domain discrepancy (combining MMD, cross-entropy, and adversarial loss terms), SS encodes a “reverse simulation” of the domain adaptation process and learns a weight matrix over source domains, modulating their contribution to the subsequent adaptation. Unlike instance-level similarity ranking, this mechanism is inherently dynamic and adapts with model training, capturing emergent relationships inaccessible to fixed similarity metrics.
The AS module consumes the reweighted source data and enforces both domain-confusion (via a GRL placed on the domain discriminator branch) and emotion class-separability (via standard cross-entropy on the emotion branch). Critically, the minimum domain classifier discrepancy (MDC) objective ensures the adversarial game between domain classifiers retains sufficient signal for robust adaptation, sidestepping the collapse of domain discernibility that can impair adversarial DA.
Augmentation is further achieved through Gaussian noise injection in feature space, enhancing robustness to input perturbations and modeling intrinsic uncertainties in affective state. Spherical Logistic Regression is employed for decision boundaries in normalized spaces, improving alignment in high-dimensional non-Euclidean EEG manifolds.
Empirical Evaluation
Comprehensive experiments on SEED, SEED-IV, and HBUED—spanning both three-class and four-class emotion recognition scenarios—demonstrate that SSAS achieves SOTA accuracy under leave-one-subject-out cross-validation (LOSOCV). Specifically, SSAS yields 91.97% on SEED and 77.99% on SEED-IV, outperforming strong adversarial and graph neural baselines by margins of up to 7% on the more heterogeneous SEED-IV. Notably, SSAS delivers robust performance improvements in low-sample and high-heterogeneity regimes, aligning with its design objective of emphasizing beneficial sources while suppressing detrimental ones.
Ablation studies highlight the indispensable role of both MMD (penalizing inter-subject discrepancy) and the MDC (preserving adversarial balance); removal of either reduces generalization by over 10%. The SS module’s contribution is amplified in datasets where negative transfer is prevalent, supporting its utility as a filter in realistic, multi-source adaptation environments. Visualization using t-SNE further validates that the post-adaptation feature clusters are better aligned with emotion classes and domains, with significant reduction in class confusion compared to vanilla DA.
Theoretical Analysis
SSAS is anchored in a formalized extension of classic DA theory, introducing an error bound over the target expected loss that incorporates both inter-domain discrepancy and a data-driven supremum over source domains selected by dynamic weighting. Unlike standard analysis that only considers average-case source alignment, SSAS operationalizes a bound that explicitly minimizes transfer from high-divergence or low-similarity sources. This is formalized in a multi-source version of the Ben-David et al. domain adaptation bound, where maximization of worst-case pairwise divergence tightens control over target risk, especially under high source heterogeneity.
Implications and Future Directions
The SSAS architecture demonstrates that source domain selection is crucial in robust cross-subject EEG emotion recognition. The explicit adversarial simulation and dynamic weighting distinguish SSAS from naïve alignment or simple instance filtering approaches, providing a blueprint for negative transfer mitigation in generic multi-source domain adaptation. In practical terms, SSAS enables higher-fidelity aBCIs deployable in settings where annotated target data is absent and inter-subject variability is high, such as real-world emotion monitoring, clinical diagnostics, and longitudinal neuroadaptive interfaces.
However, SSAS incurs computational overhead for explicit source weighting and sequential training; its marginal advantage is attenuated in settings where sources are uniformly high-quality. Future work is likely to explore lightweight or continual selection modules, soft adaptive gating, and integration with more sophisticated uncertainty quantification or generative data augmentation for improved sample efficiency.
Conclusion
SSAS establishes an empirically validated and theoretically motivated framework for source selection and adversarial domain adaptation in cross-subject EEG emotion recognition. By coupling dynamic, adversarially informed weighting with robust adaptation objectives, it attains high generalization accuracy, particularly in challenging, negative transfer-prone environments. Extensions of SSAS are likely to inform broader DA practice for other bio-signal recognition and multi-source adaptation tasks.