FACE: Few-shot Adapter with Cross-view Fusion for Cross-subject EEG Emotion Recognition
Authors: Haiqi Liu, C. L. Philip Chen, Tong Zhang
Organizations: the Guangdong Provincial Key Laboratory of Computational AI Models and Cognitive Intelligence, the School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China · the Pazhou Lab, Guangzhou 510335, China · Engineering Research Center of the Ministry of Education on Health Intelligent Perception and Paralleled Digital-Human, Guangzhou, China
Cross-subject EEG emotion recognition is challenged by significant inter-subject variability and intricately entangled intra-subject variability. Existing works have primarily addressed these challenges through domain adaptation or generalization strategies. However, they typically require extensive target subject data or demonstrate limited generalization performance to unseen subjects. Recent few-shot learning paradigms attempt to address these limitations but often encounter catastrophic overfitting during subject-specific adaptation with limited samples. This article introduces the few-shot adapter with a cross-view fusion method called FACE for cross-subject EEG emotion recognition, which leverages dynamic cross-view fusion and effective subject-specific adaptation. Specifically, FACE incorporates a cross-view fusion module that dynamically integrates global brain connectivity with localized patterns via subject-specific fusion weights to provide complementary emotional information. Moreover, the few-shot adapter module is proposed to enable rapid adaptation for unseen subjects while reducing overfitting by enhancing adapter structures with meta-learning. Experimental results on three public EEG emotion recognition benchmarks demonstrate the effectiveness of FACE under few-shot cross-subject evaluation.
Figures & tables
Fig. 1: Comparison of training data and processes between Few-Shot Learning (FSL) and traditional deep learning (DL) in cross-subject EEG emotion recognition. Traditional DL primarily relies on additional unlabeled target data for training, whereas FSL learns a subject-agnostic model in the source domain and fine-tunes it using only a few target-domain samples with few steps. DA denotes Domain Adaptation, DG denotes Domain Generalization, and SSL denotes Semi-Supervised Learning.
Fig. 2: Overview of the proposed FACE architecture in the testing stage. For each unseen subject, a few labeled EEG signals are collected for rapid partial updates. Their EEG signals are spatially projected to spatial representations, which are then fed into FACE. The framework contains two core components: 1) CVF learns subject-specific fusion of global graph and local spatial features, and 2) The Few-Shot Adapter Module adjusts the features to ensure subject-specific emotional cues are captured.
Fig. 3: Illustration of the FACE training process: (a) inner-loop subject adaptation on the support set, and (b) outer-loop meta-update using query losses across n subjects.
Dataset
Num of Subjects
Data per Subject per Session
Official Length
DE Samples per Subject per Session
SEED
15
5 trials × 3 classes
1s
3394
SEED-IV
15
6 trials × 4 classes
4s
822 ∼ 851
SEED-V
16
3 trials × 5 classes
4s
541 ∼ 681
TABLE I: Statistics of the datasets used in this study.
Method
#Num of unlabeled target samples
#Num of labeled target samples
SEED Mean / STD (%)
SEED-IV Mean / STD (%)
Supervised Learning
IAG [ 26 ]
✗
✗
86.30 / 6.91
62.64 / 10.25
GMSS [ 20 ]
86.5 / 6.22
73.48 / 7.41
PGCN [ 19 ]
84.59 / 8.68
73.69 / 7.16
Semi-Supervised learning
SDDA [ 8 ]
ALL
∼ 225/ ∼ 35
91.08 / 7.70
81.58 / 8.72
SSML [ 29 ]
10-shot
88.59
-
TABLE II: Comparison with state-of-the-art approaches on SEED and SEED-IV. The best results are shown in bold, and the second-best results are underlined.
Method
#NUS
#NLS
SEED-V Mean / STD (%)
IAG [ 26 ]
✗
✗
59.68 / 9.44
Progressive GCN [ 56 ]
71.40 / 9.43
PGCN ∗ [ 19 ]
65.21 / 10.10
JAGP [ 57 ]
ALL
✗
75.43 / 7.57
SSAA-MSDA ∗ [ 53 ]
63.09 / 10.01
PR-PL ∗ [ 54 ]
72.62 / 12.67
TABLE III: Comparison with state-of-the-art approaches on SEED-V. The best results are shown in bold, and the second-best results are underlined.
Fig. 4: Statistical comparison of different methods on the (a) SEED, (b) SEED-IV, and (c) SEED-V datasets. The asterisks (*) indicate 0.0001<p≤0.001 , and four asterisks (**) denote p≤0.0001 , with more asterisks indicating stronger statistical significance. “ns” indicates no significant difference.
CVF
FSA
SEED
SEED-IV
SEED-V
3-shot
5-shot
10-shot
3-shot
5-shot
10-shot
3-shot
5-shot
10-shot
81.80 ± 7.79
86.60 ± 6.27
89.70 ± 5.36
80.37 ± 4.33
84.65 ± 3.75
88.21 ± 3.07
78.31 ± 5.23
82.63 ± 4.70
85.20 ± 5.17
✓
90.84 ± 5.26
93.70 ± 3.85
95.36 ± 3.03
82.57 ± 4.44
87.03 ± 3.86
90.73 ± 3.26
87.60 ± 3.80
91.08 ± 3.38
94.63 ± 2.23
✓
84.51 ± 7.04
88.86 ± 5.48
94.02 ± 3.77
83.48 ± 4.16
88.75 ± 2.74
95.10 ± 1.30
87.81 ± 2.81
94.09 ± 1.58
98.23 ± 0.87
✓
✓
91.66 ± 3.72
93.96 ± 2.70
96.72 ± 1.81
83.91 ± 3.87
89.51 ± 3.13
95.95 ± 1.47
89.55 ± 2.98
95.20 ± 1.42
98.95 ± 0.36
TABLE IV: Ablation studies of the proposed method on the SEED, SEED-IV, and SEED-V datasets, with the best results highlighted in bold.
Fig. 5: Statistical comparison of different components on the (a) SEED, (b) SEED-IV, and (c) SEED-V datasets. The asterisk ( ∗ ) indicates 0.01\textlessp≤0.05 while the four asterisks (****) denote p≤0.0001 , with a higher number of asterisks representing a more significant improvement.
SEED
SEED-IV
SEED-V
DGCN
86.60 ± 6.27
84.65 ± 3.75
82.63 ± 4.70
Conv3
88.85 ± 5.39
80.74 ± 6.32
90.53 ± 4.73
CVF (DGCN+Conv3)
93.70 ± 3.85
87.03 ± 3.86
91.08 ± 3.38
DGCNN
83.60 ± 6.73
62.75 ± 6.96
69.55 ± 11.75
Conv3
88.04 ± 4.88
79.63 ± 5.81
90.41 ± 4.64
CVF (DGCNN+Conv3)
88.55 ± 5.91
67.47 ± 7.00
87.63 ± 6.14
TABLE V: Comparison of the CVF Module with Different Backbones on SEED, SEED-IV, and SEED-V
SEED
SEED-IV
SEED-V
Self Attention
78.38 ± 7.91
70.59 ± 7.04
65.62 ± 8.29
Cross Attention
71.91 ± 9.83
59.69 ± 4.98
73.17 ± 9.70
CVF w/o MAML
93.28 ± 4.10
86.58 ± 3.82
90.52 ± 3.42
CVFboth
92.93 ± 3.75
86.15 ± 3.40
90.79 ± 3.84
CVF(ours)
93.70 ± 3.85
87.03 ± 3.86
91.08 ± 3.38
TABLE VI: Comparison of the CVF Module with Different Fusion Strategies on SEED, SEED-IV, and SEED-V
Fig. 6: t-SNE visualization of the SEED and SEED-IV datasets with and without the CVF module under the 5-shot setting. (a) Baseline w/o CVF module on SEED, (b) Baseline w/ CVF module on SEED, (c) Baseline w/o CVF module on SEED-IV, and (d) Baseline w/ CVF module on SEED-IV. Markers represent the ground truth labels, while colors indicate the predicted labels.
SEED
SEED-IV
SEED-V
Linear Probing
84.40 ± 6.57
84.24 ± 4.03
83.64 ± 4.45
LoRA
87.55 ± 3.78
83.14 ± 3.57
79.29 ± 5.62
Serial Adapter
87.58 ± 5.87
85.11 ± 4.13
84.75 ± 5.44
Residual Adapter
88.09 ± 5.83
86.96 ± 3.46
88.06 ± 3.46
FSA(ours)
88.86 ± 5.48
88.75 ± 2.74
94.09 ± 1.58
TABLE VII: Comparison of the FSA Module with Different Parameter-efficient Methods on SEED, SEED-IV, and SEED-V
Fig. 7: Brain topographic maps illustrating the effects of 5-shot adaptation on (a) the SEED and (b) the SEED-IV dataset. The first row depicts the state before adaptation, while the second row represents the state after adaptation.
Fig. 8: 3-shot and 5-shot results on SEED, SEED-IV, and SEED-V. Setting 1 draws adaptation samples from one trial per class and tests on the other trials. Setting 2 retains about half of the trials, Setting 3 uses the standard protocol, and Setting 4 uses all three sessions of the source subjects for training.
Fig. 9: Accuracy Trends of SEED, SEED-IV and SEED-V across Different Shot Numbers
Method
SEED-IV (1s)
SEED-IV (4s)
SEED-V (1s)
SEED-V (4s)
PR-PL [ 54 ]
81.32 ± 8.53
–
–
–
PR-PLavg∗
71.96 ± 7.56
74.46 ± 4.56
64.19 ± 8.48
72.62 ± 12.67
PR-PLmax∗
81.89 ± 9.57
84.64 ± 4.49
76.30 ± 11.17
83.52 ± 11.17
SSAA-MSDA [ 53 ]
–
76.23 ± 9.02
–
–
SSAA-MSDAavg∗
67.91±6.47
73.65 ± 6.73
53.69±6.09
63.09 ± 10.01
SSAA-MSDAmax∗
78.40±10.34
83.73 ± 7.64
62.75±5.43
72.57 ± 10.55
TABLE VIII: Classification accuracy (%) on SEED-IV and SEED-V datasets under 1s and 4s segment lengths.
Cross-subject generalization remains a major challenge for EEG-based emotion recognition due to pronounced inter-subject variability in neural activity and emotion-related spatiotemporal patterns. Existing methods mainly alleviate inter subject differences through distribution alignment or contrastive learning, while paying limited attention to representation bias associated with fixed spatiotemporal modeling orders. We hypothesize that high-level emotion representations of the same EEG signal should remain consistent across different spatiotemporal modeling paths. Based on this hypothesis, we propose a Dual-Path Spatiotemporal Contrastive Learning (DP-STCL) framework for cross-subject EEG emotion recognition. DP-STCL employs two complementary encoders with independent parameters, namely a Temporal-to-Spatial (TS) path and a Spatial to-Temporal (ST) path, to model EEG spatiotemporal information under different integration orders. During representation learning, the same emotional stimulus is used to establish cross subject associations, while three complementary relations jointly constrain representation variations across subjects and modeling paths. This promotes stable and discriminative stimulus-related emotion representations. To further handle limited labeled target subject data, we introduce a LoRA-based few-shot adaptation strategy with low parameter overhead. Extensive cross-subject experiments on SEED, SEED-IV, SEED-V, and FACED show that DP-STCL consistently outperforms the compared methods, demonstrating its effectiveness in improving cross-subject EEG emotion recognition.
Zehui Xiao, Yi Zheng, Guixing Fu +7
School of Biomedical Engineering, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China · School of Computer Science and Technology, Tianjin University, Tianjin 300354, China · Faculty of Applied Sciences, Macao Polytechnic University, Macao, China +1
Electroencephalogram (EEG) captures endogenous brain activity with high temporal fidelity and holds substantial promise for precise emotion decoding. However, channel redundancy and pronounced inter-subject variability remain key obstacles to scalable generalization. To address these limitations, we propose a novel framework termed PRioritized channel Importance with Semi-supervised doMain adaptation (PRISM), enabling label-efficient cross-subject emotion decoding. On the channel side, PRISM assigns differentiable, data-dependent channel weights via a lightweight expert ensemble, amplifying reliable electrodes while suppressing distractors. On the domain side, PRISM leverages unlabeled data through confidence-filtered pseudo-labels to drive consistency regularization and domain alignment, mitigating subject-specific heterogeneity. Extensive experiments show that PRISM surpasses state-of-the-art methods on DEAP, DREAMER, and SEED datasets, achieving robust cross-subject generalization given limited annotations.
Xin Zhou, Xiang Zhang, Hao Deng +1
School of Computing, T. J Watson College of Engineering and Applied Science, Binghamton University - State University of New York, Binghamton, NY 13902 USA · Massachusetts General Hospital, Harvard University, Boston, MA 02114 USA
With the advancement of science and technology, the importance of emotion research has become increasingly evident. Electroencephalography (EEG)-based emotion recognition has emerged as an active research area in recent years, owing to its objectivity and high temporal resolution. However, most existing methods focus on optimizing encoder structures to enhance feature extraction capabilities, while paying relatively little attention to similarity calculation strategies, particularly overlooking the potential temporal misalignment of responses among different subjects. To address these shortcomings, this paper draws inspiration from the late interaction mechanism of ColBERT in natural language processing (NLP) and proposes a Temporal Asynchronous Alignment-based Contrastive Learning (TA2CL) framework. This method transforms the traditional global "hard alignment" similarity calculation approach into a fine-grained local matching mechanism, enabling the model to adaptively search for and align "locally highly correlated" segments between two EEG signals, thereby effectively mitigating the effects of inter-subject differences and temporal delays. Experimental results demonstrate that the proposed method achieves strong performance across multiple public datasets. Specifically, on the FACED dataset, it achieves an accuracy of 64.5% for the nine-class classification task and 79.5% for the binary classification task, while on the SEED and SEED-V datasets, it achieves accuracies of 86.4% and 70.1%, respectively, validating the method's effectiveness and generalization capability.
Ying Xie, Yi Zheng, Zehui Xiao +2
School of Biomedical Engineering, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China · School of Computer Science and Technology, Tianjin University, Tianjin 300354, China