Abstract：Human emotions are closely linked to specific brain regions. Recent studies have increasingly focused on extracting regional representations for EEG-based emotion recognition (EER). However, the EEG signals associated with these regions exhibit high individual specificity in both feature distributions and brain connectivity patterns, making effective generalization for cross-subject EER a key challenge. To address this challenge, we proposed an adaptive mixture of experts (AdaMoE), leveraging MoE's expert specialization mechanism for specialized regional processing while introducing a dual test-time adaptation strategy over feature representations and graph topology to enhance cross-subject generalizability. Specifically, AdaMoE utilized multiple graph convolution experts that adaptively process different combinations of EEG channels and their corresponding inter-regional interaction patterns to capture diverse regional representations. On this basis, feature adaptation dynamically calibrated target sample representations to improve the domain-invariance of emotion-discriminative features, while topology adaptation adaptively adjusts graph connections to more accurately characterize subject-specific brain-region interaction patterns. Extensive experiments on two public datasets demonstrated that AdaMoE achieved accuracies of 80.5% and 57.7%, with F1 scores of 79.6% and 56.5% on the SEED and SEED-IV datasets, respectively, thereby effectively improving cross-subject EEG emotion recognition performance. These results demonstrate that regional graph-expert modeling and dual feature-topology test-time adaptation can jointly mitigate feature and structural shifts in unseen subjects while maintaining stable online updates.
[1] Scherer KR, Moors A. The emotion process: event appraisal and component differentiation[J]. Annu Rev Psychol, 2019, 70(1): 719-745. [2] D Mello S, Kory J. A review and meta-analysis of multimodal affect detection systems[J]. ACM Comput Surv, 2015, 47(3): 1-36. [3] Wu D, Lu BL, Hu B, et al. Affective brain-computer interfaces: a tutorial[J]. Proc IEEE, 2023, 111(10): 1314-1332. [4] Alarcao SM, Fonseca MJ. Emotions recognition using EEG signals: a survey[J]. IEEE Trans Affect Comput, 2017, 10(3): 374-393. [5] Craik A, He Y, Contreras-Vidal JL. Deep learning for electroencephalogram (EEG) classification tasks: a review[J]. IEEE J Biomed Health Inform, 2019, 23(3): 1482-1490. [6] Li X, Zhang Y, Tiwari P, et al. EEG based emotion recognition: a tutorial and review[J]. ACM Comput Surv, 2022, 55(4): 1-57. [7] 周哲昊, 孙托, 吴灿标, 等. 脑电情绪识别模型研究进展[J]. 湘潭大学学报(自然科学版), 2025, 47(3): 35-53. [8] 陈旭东, 钟恒, 皇甫洁, 等. 脑电信号情绪识别综述[J]. 计算机应用, 2023, 43(S1): 323-332. [9] 杨江江, 乔晓艳. 多源域子域自适应网络脑电信号情感识别[J]. 网络新媒体技术, 2024, 13(5): 26-33. [10] Li F, Zhang D. Multimodal physiological signals from wearable sensors for affective computing: a systematic review[J]. Intelligent Sports and Health, 2025, 1(4): 210-222. [11] Cacioppo JT, Tassinary LG, Berntson GG. Handbook of Psychophysiology[M](The 3rd edition). Cambridge: Cambridge University Press, 2007. [12] Zeng Y, Zhao Q, Wang Y, et al. Review of physiological-signal-based emotion recognition[J]. J Biomed Eng, 2021, 38(5): 905-918. [13] Koelstra S, Muhl C, Soleymani M, et al. DEAP: a database for emotion analysis using physiological signals[J]. IEEE Trans Affect Comput, 2012, 3(1): 18-31. [14] Niedermeyer E, da Silva FL. Electroencephalography: Basic Principles, Clinical Applications, and Related Fields[M] (The 6th edion). Philadelphia: Lippincott Williams & Wilkins, 2011. [15] Hu B, Li X, Sun S, et al. THU-EP: a large-scale benchmark for cross-subject EEG-based emotion prediction[J]. IEEE J Biomed Health Inform, 2022, 26(5): 2141-2152. [16] Li W, Huan W, Hou B, et al. Can emotion be transferred?—a review on transfer learning for EEG-based emotion recognition[J]. IEEE Trans Cogn Dev Syst, 2022, 14(3): 833-846. [17] Chen H, Jin M, Li ZN, et al. MS-MDA: multisource marginal distribution adaptation for cross-subject and cross-session EEG emotion recognition[J]. Front Neurosci, 2021, 15: 778488. [18] Zhong P, Wang D, Miao C. EEG-based emotion recognition using regularized graph neural networks[J]. IEEE Trans Affect Comput, 2020, 13(3): 1290-1301. [19] Ma B Q, Li H, Zheng W L, et al. Reducing the subject variability of EEG signals with adversarial domain generalization[C]//International Conference on Neural Information Processing. Cham: Springer, 2019: 30-42. [20] Shen XK, Liu XG, Hu X, et al. Contrastive learning of subject-invariant EEG representations for cross-subject emotion recognition[J]. IEEE Trans Affect Comput, 2022, 14(3): 2496-2511. [21] Cai HH, Pan JH. Two-phase prototypical contrastive domain generalization for cross-subject EEG-based emotion recognition[C]//2023 IEEE International Conference on Acoustics, Speech and Signal Processing. Piscataway: IEEE, 2023: 1-5. [22] Li Y, Zhang H, Chen X, et al. NPICL: neuron-inspired parallel contrastive learning for cross-subject EEG emotion recognition[J]. Pattern Recognit, 2025, 158: 111042. [23] Ding Y, Tong CX, Zhang SL, et al. EmT: a novel transformer for generalized cross-subject EEG emotion recognition[J]. IEEE Trans Neural Netw Learn Syst, 2025, 36(6): 10381-10393. [24] Liu XH, Jiang WB, Zheng WL, et al. MoGE: mixture of graph experts for cross-subject emotion recognition via decomposing EEG[C]//2024 IEEE International Conference on Bioinformatics and Biomedicine. Piscataway: IEEE, 2024: 3515-3520. [25] Imtiaz MN, Khan N. Enhanced cross-dataset electroencephalogram-based emotion recognition using unsupervised domain adaptation[J]. Comput Biol Med, 2025, 184: 109394. [26] Shao H, Dong L, Li L, et al. PFC-TTA: parameter-free classifier for test time adaptation in emotion recognition[C]//IEEE International Joint Conference on Neural Networks. Rome: IEEE, 2025: 1-8. [27] Zheng WL, Zhu JY, Lu BL. Identifying stable patterns over time for emotion recognition from EEG[J]. IEEE Trans Affect Comput, 2019, 10(3): 417-429. [28] Zheng WL, Liu W, Lu YF, et al. EmotionMeter: a multimodal framework for recognizing human emotions[J]. IEEE Trans Cybern, 2019, 49(3): 1110-1122. [29] Duan RN, Zhu JY, Lu BL. Differential entropy feature for EEG-based emotion classification[C]//The 6th International IEEE EMBS Conference on Neural Engineering. San Diego: IEEE, 2013: 81-84. [30] Shi LC, Jiao YY, Lu BL. Differential entropy feature for robust emotion recognition from EEG signals[C]//The 7th International IEEE EMBS Conference on Neural Engineering. Cancun: IEEE, 2015: 917-920. [31] Gao Z, Wang X, Yang Y, et al. A channel-fused dense convolutional network for EEG-based emotion recognition[J]. IEEE Trans Cogn Dev Syst, 2020, 13(4): 945-954. [32] Zhang T, Wang X, Xu X, et al. GCB-Net: graph convolutional broad network for EEG emotion recognition[J]. IEEE Trans Affect Comput, 2021, 12(2): 544-556. [33] He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//Proc IEEE Conf Comput Vis Pattern Recognit. Las Vegas: IEEE, 2016: 770-778. [34] Wang DQ, Shelhamer E, Liu ST, et al. Tent: fully test-time adaptation by entropy minimization[C/OL]//International Conference on Learning Representations. https://openreview.net/forum?id=uXl3bZLkr3c. 2021-03-18/2026-04-01. [35] Wang Q, Fink O, Van Gool L, et al. CoTTA: continual test-time domain adaptation[C]//Proc IEEE/CVF Conf Comput Vis Pattern Recognit. New Orleans: IEEE, 2022: 11964-11974.
