A Multi-Source Domain EEG Emotion Recognition Method Based on Self-Organizing Sparse Directed Graph Convolution
By adaptively modeling the directional relationships of EEG channels using a self-organizing sparse directed graph convolutional network, the problems of cross-subject EEG signal distribution differences and multi-source domain instability were solved, achieving higher recognition accuracy and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU DIANZI UNIV
- Filing Date
- 2026-02-14
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies are unable to effectively address the issues of large differences in the distribution of EEG signals across subjects, insufficient modeling of channel topology relationships, and unstable multi-source domain migration, resulting in insufficient accuracy and stability of cross-subject EEG emotion recognition models.
We employ a self-organizing sparse directed graph convolutional network to capture the asymmetric information propagation characteristics between EEG channels by constructing a sparse adjacency matrix and a direction-sensitive message passing matrix. Furthermore, we perform collaborative alignment and dynamic weighting in multi-source domain scenarios to construct domain-invariant EEG emotional features.
It significantly improves the accuracy and stability of cross-subject EEG emotion recognition, and achieves stronger model generalization ability and recognition performance by adaptively modeling the directional relationship of EEG channels.
Smart Images
Figure CN122087422A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to methods for electroencephalogram (EEG) signal processing and emotion recognition, and particularly to a multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution. Background Technology
[0002] Emotional state is a crucial manifestation of higher cognitive activities and psychological states in humans, closely related to an individual's decision-making behavior, learning efficiency, and mental health level. Objective and accurate identification of individual emotional states has significant application value in fields such as affective computing, human-computer interaction, brain-computer interfaces, intelligent healthcare, and mental health assessment. Compared to overt behavioral signals such as facial expressions and speech, electroencephalogram (EEG) signals directly reflect brain neural activity, offering advantages such as high temporal resolution and resistance to spoofing, making them a vital data source for emotion recognition research.
[0003] However, EEG signals exhibit significant non-stationarity and individual variability. Different subjects show marked differences in brain structure, neural response patterns, physiological states, and experimental environments. Even under the same emotional state, their EEG signals may still exhibit significantly different distribution characteristics in the time domain, frequency domain, and spatial channel connectivity patterns. This cross-subject distribution shift problem makes it difficult to directly transfer emotion recognition models trained on a single subject or a single data source to new subjects, thus severely limiting the practical application of cross-subject EEG emotion recognition.
[0004] To alleviate these problems, researchers have introduced transfer learning, especially domain adaptation methods, to improve the generalization ability of models by aligning the feature distributions of the source and target domains. Existing methods mostly model alignment from the perspective of overall feature distribution or conditional distribution, which improves cross-subject recognition performance to some extent. However, these methods still have the following shortcomings: First, most methods ignore the inherent spatial topology and potential directional information between EEG channels, making it difficult to characterize the asymmetric information transmission relationships between brain regions; second, traditional graph convolutional networks typically use undirected or predefined fixed topologies, lacking the ability to adaptively model EEG channel relationships; third, in multi-source domain transfer scenarios, the correlation between different source domains and the target domain varies. If the reliability of transfer between each source domain is not effectively evaluated, weak source domain interference can easily be introduced, leading to negative transfer phenomena.
[0005] Therefore, there is an urgent need for a cross-subject EEG emotion recognition method that can adaptively model the directional relationship of EEG channels and achieve collaborative alignment and dynamic weighting in multi-source domain scenarios, so as to improve the accuracy, stability and generalization ability of the model. Summary of the Invention
[0006] To address the aforementioned technical deficiencies, this invention proposes a multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution, which solves the problems of large differences in cross-subject EEG signal distribution, insufficient modeling of channel topology relationships, and unstable multi-source domain migration in the aforementioned background technologies.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution includes the following steps: S1. Acquire and preprocess EEG signals; construct EEG feature vectors based on EEG signals; S2. Construct a sparse adjacency matrix based on EEG feature tensors; S3. Perform directed graph convolution operations on the forward and reverse adjacency matrices of the sparse adjacency matrix respectively to capture the asymmetric information propagation characteristics between EEG channels and generate domain-invariant EEG emotional features. S4. For multiple source domain samples and target domain samples, construct multiple domain feature extraction branches; input domain-invariant EEG emotion features into the multiple domain feature extraction branches respectively; construct independent feature representation spaces for each source domain and target domain, and separate domain-shared features and domain-specific features; classifier classifies samples based on domain-shared features and domain-specific features; S5. Integrate the prediction results of each source domain classifier for the target domain samples and output the EEG emotion recognition result of the target domain; use the weighted result of the overall loss function of each source domain to perform cross-domain optimization of the model.
[0008] Preferably, S1 includes: The acquired raw EEG signals were bandpass filtered and divided into frequency bands according to different datasets. The signal variance of each EEG channel in each frequency band was calculated within a non-overlapping time window, and the differential entropy features of the EEG signals were extracted based on the differential entropy theory. The differential entropy features were then subjected to time smoothing using a linear dynamic system to obtain the EEG feature tensor.
[0009] Preferably, S2 includes: S21. Map the EEG feature tensor to the latent space to generate the mapped feature matrix; S22. Construct an adjacency matrix based on the mapped feature matrix to represent the adjacency relationship between channels; S23. Construct a sparse adjacency matrix based on the adjacency matrix.
[0010] As a preferred embodiment, S23 includes: A Top-k sparsity strategy is applied to the adjacency matrix, retaining only the channel with the largest connection weight. Connect 1, and reset the weights of the remaining connections to zero, constructing a sparse adjacency matrix: in, This represents Hadamard element-wise multiplication. Let A represent the sparse mask matrix, and let A be the adjacency matrix.
[0011] Preferably, S3 includes: S31. Based on the sparse adjacency matrix, define the forward adjacency matrix and the reverse adjacency matrix, and then construct the direction-sensitive message passing matrix. S32. Feature extraction of the message passing matrix is performed at multiple scales through forward propagation, backward propagation, and self-loop information. Finally, the multi-scale features are fused through an attention mechanism to obtain domain-invariant EEG emotion features.
[0012] Preferably, in S32, the feature update process is represented as follows: in, It is the first Layer output; , , These are the forward weight matrix, the reverse weight matrix, and the self-loop weight matrix, respectively. It is the fusion coefficient, used to balance the forward and reverse information flows; It represents the characteristic information propagating from the source node to the target node, reflecting the strength of the forward flow of the signal; It indicates the reverse flow of information.
[0013] As a preferred embodiment, S5 includes: S51. Construct the overall loss function; S52. Based on the prediction results of each source domain classifier for the target domain samples, the prediction confidence is obtained: the average of each prediction confidence is taken as the source domain transfer performance index; the source domain weight coefficient is obtained by using the exponential amplification function based on the source domain transfer performance index. S53. The overall loss function of each source domain branch is weighted according to the source domain weight coefficient to achieve cross-domain optimization; the prediction results of each source domain are fused by weighting the source domain weight coefficient to output the EEG emotion recognition result of the target domain.
[0014] Preferably, S51 includes: Construct a multi-branch consistency constraint loss to constrain the prediction consistency of the target domain across different source domain branches: in, The number of source domain classifiers, For the first The predicted vectors of each branch in the target domain.
[0015] As a preferred embodiment, in S5, the overall loss function is: in, , , For loss weight hyperparameters, For source domain supervised classification loss, The edge distribution alignment loss is based on the normalized Wasserstein distance. This is the conditional distribution alignment loss based on the local maximum mean difference. The loss is due to the multi-branch consistency constraint.
[0016] Compared with the prior art, the beneficial effects of the present invention are reflected in: 1. Unlike traditional techniques that use undirected graphs or predefined fixed topologies to model EEG channel relationships, this invention proposes a self-organizing sparse directed graph convolutional network. This network can dynamically learn the directional topological connections between EEG channels from data and adaptively capture the asymmetric propagation characteristics of neural signals through forward and backward information flow modeling. This enables the model to construct a physiological model that more closely matches the flow of neural information in the brain, thereby significantly improving the discriminative representation ability of emotion-related EEG features.
[0017] 2. Unlike traditional methods that often employ a single distribution alignment strategy, this invention constructs a multi-source domain adaptive framework integrating multi-level alignment and consistency constraints. By integrating edge distribution alignment, conditional distribution alignment, local maximum mean difference, and cross-branch consistency constraints, it achieves more robust cross-subject EEG feature transfer than a single alignment mechanism. This enables the model to achieve more robust and discriminative cross-subject EEG feature transfer than a single alignment mechanism.
[0018] 3. Unlike existing methods that treat multiple source domains equally or employ fixed weighting strategies, this invention designs a performance-driven source domain weighting mechanism (PDWM). This mechanism dynamically evaluates and quantifies the transfer reliability of different source domains based on the prediction confidence of each source domain classifier on the target domain samples, and adaptively adjusts their weights in training and decision-making. This enables the model to effectively suppress the "negative transfer" effect caused by weakly correlated or interfering source domains, significantly improving the generalization ability and stability of the cross-subject emotion recognition model. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the method principle in Embodiment 1 of the present invention; Figure 2 This is a channel weighting diagram between the upper electrodes of the SEED in Embodiment 1 of the present invention; Figure 3This is an EEG topographic map on the SEED in Embodiment 1 of the present invention. Detailed Implementation
[0020] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.
[0021] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.
[0022] This invention uses EEG data as input and obtains stable EEG feature representations through differential entropy feature extraction and linear dynamic system smoothing. It further constructs a self-organizing sparse directed graph convolutional network to adaptively model the directional dependencies between EEG channels. Within a multi-source domain collaborative learning framework, it introduces various loss functions such as edge distribution alignment, conditional distribution alignment, and multi-branch consistency constraints to achieve effective alignment between the source and target domains in a multi-level feature space. Simultaneously, based on a performance-driven source domain weighting mechanism, it dynamically adjusts the transfer contributions of different source domains, thereby suppressing weak source domain interference and improving the accuracy and stability of cross-subject EEG emotion recognition.
[0023] Example 1: like Figure 1 The method for multi-source domain EEG emotion recognition based on self-organizing sparse directed graph convolution, as shown, includes the following steps: S1. Acquire and preprocess EEG signals; construct EEG feature vectors based on EEG signals; S11. Bandpass filtering is performed on the acquired multi-channel EEG signals to divide the EEG signals into multiple frequency bands. (1-3 Hz), (4-7 Hz), (8-13 Hz), (14-30 Hz), (31-50 Hz), and calculate the signal variance of each EEG channel in each frequency band within a preset time window; S12, since the original EEG signal can be approximated as a stationary Gaussian process after bandpass filtering, calculate the differential entropy characteristics of the EEG signal according to the definition of differential entropy, and the calculation formula is as follows: in, Represents a random variable of electroencephalogram (EEG) signals. This represents the variance of the EEG signal within the corresponding time window. This represents the mean value of the EEG signal within the corresponding time window.
[0024] S13. Further perform time smoothing on the differential entropy features to construct an EEG feature tensor for subsequent modeling.
[0025] S2. Constructing a self-organizing sparse directed graph based on EEG feature tensors; S21. Map the EEG feature tensor to the latent space to generate the mapped feature matrix; Based on the EEG feature tensor, let the input feature matrix be: in, Indicates the number of brainwave channels. This represents the feature dimension of each channel; The input features are mapped to the latent space through a learnable linear mapping, and the mapping process is represented as follows: in, For the mapping matrix, This represents the feature matrix after mapping. The activation function ensures that the feature map has a non-linear compression effect and suppresses extreme value interference.
[0026] S22. Construct an adjacency matrix based on the mapped feature matrix to represent the adjacency relationship between channels; Based on the mapped feature matrix, the adjacency relationship between channels is constructed using the node inner product method. The formula for calculating the adjacency matrix is as follows: in Represents a node Vectors in the mapping space, Indicates from node Pointing to node The probability distribution of connection strength, where N represents the number of EEG channels.
[0027] S23. Construct a sparse adjacency matrix based on the adjacency matrix; Furthermore, a Top-k sparsity strategy is applied to the adjacency matrix, retaining only the channel with the largest connection weight. Connect 1, and reset the weights of the remaining connections to zero, constructing a sparse adjacency matrix: in This represents Hadamard element-wise multiplication. This represents a sparse mask matrix.
[0028] S3. Perform directed graph convolution operations on the forward and reverse adjacency matrices of the sparse adjacency matrix respectively to capture the asymmetric information propagation characteristics between EEG channels and generate domain-invariant EEG emotional features. S31. Define the degree matrix based on the sparse adjacency matrix. and in-degree matrix This allows for the construction of a direction-sensitive message passing matrix. Out-degree matrix and in-degree matrix The definition is as follows: This leads to the construction of a direction-sensitive message passing matrix: in It represents the characteristic information propagating from the source node to the target node, reflecting the strength of the forward flow of the signal; This indicates reverse information flow.
[0029] S32. Multi-scale feature extraction is performed on the message passing matrix through forward propagation, backward propagation, and self-loop information, and the multi-scale features are fused to obtain domain-invariant EEG emotion features.
[0030] In the process of directed graph convolution, forward propagation, backward propagation, and self-loop information are introduced, and its feature update process is represented as follows: in It is the first Layer output; , , These are the forward, reverse, and self-loop weight matrices, respectively. It is the fusion coefficient, used to balance the forward and reverse information flows.
[0031] Multi-branch feature extraction is performed on the above directed graph convolution process using a multi-scale parallel structure, and the multi-scale features are fused to obtain domain-invariant EEG emotion features.
[0032] S4. Based on domain-invariant EEG emotion features, separate domain-shared features and domain-specific features; classifier classifies samples based on domain-shared features and domain-specific features; Domain-invariant EEG emotional features are input into multiple domain-specific feature extraction branches to construct independent feature representation spaces for each source and target domain, thereby achieving the separation of domain-shared features from domain-specific features.
[0033] S5. Integrate the prediction results of each source domain classifier for the target domain samples and output the EEG emotion recognition result of the target domain; use the weighted result of the overall loss function of each source domain to perform cross-domain optimization of the model.
[0034] S51. Construct the overall loss function; The supervised classification loss is introduced for the source domain samples, and its calculation formula is as follows: in The number of samples in the source domain. For the number of categories, For the actual one-hot labels input, This is the model's predicted output probability for that sample, generated by the classifier after calculating the extracted features. Each source domain corresponds to an independent and unique classifier, and each classifier follows the same training strategy, using supervised classification loss to iteratively update the model parameters. To address the discrepancies in global marginal distributions, a normalized Wasserstein distance based on the kernel norm is introduced to constrain the consistency of feature distributions between the source and target domains. In adversarial learning, instead of directly training a discriminator, the prediction matrix is utilized by the classifier's output. ,in The prediction matrix representing the target domain. The prediction matrix of the source domain is represented by the loss function as follows: in The kernel norm of a matrix is used to measure the rank of a prediction matrix.
[0035] Furthermore, a subdomain alignment loss based on the local maximum mean difference is introduced to achieve conditional distribution alignment. Its calculation formula is as follows: in and Categories in the source and target domains The number of samples, the target domain estimated using pseudo-labels, For RKHS mapping functions, Represents samples from the source domain. This represents a sample from the target domain.
[0036] Finally, a multi-branch consistency constraint loss is constructed to constrain the prediction consistency of the target domain across different source domain branches: in The number of source domain classifiers, For the first The predicted vectors of each branch in the target domain.
[0037] The final overall loss of the model is: in , , This is the loss weight hyperparameter.
[0038] The source domain weighting and decision fusion steps are driven by the calculation of prediction confidence performance. S52. Based on the prediction results of each source domain classifier for the target domain samples, calculate the prediction confidence: The average confidence score of the target domain samples is used as the source domain transfer performance metric. in The target domain sample size is [number]. The source domain weight coefficients are further obtained using an exponential scaling function. in This is the magnification factor.
[0039] S53. Finally, this weight is applied to the overall loss function for each source domain branch: By incorporating weighting coefficients into the overall loss function of each source domain, performance-driven cross-domain optimization is achieved. Furthermore, by weighted fusion of the prediction results from each source domain, the EEG emotion recognition result of the target domain is output.
[0040] This invention was conducted on the SEED dataset, provided by Shanghai Jiao Tong University, which contains EEG recordings from 15 healthy university students (7 males and 8 females, with an average age of approximately 23.3 years). The experiment consisted of three independent sessions. In each session, each participant watched 15 different emotional video clips to induce three emotional states (positive, neutral, and negative). Each trial included a 5-second cue phase, approximately 4 minutes of video playback, and several rest intervals. EEG signals were acquired using a 62-channel electrode cap at a sampling frequency of 1000 Hz. The raw signals were downsampled to 200 Hz and passed through a 0–75 Hz bandpass filter to remove high-frequency noise and baseline drift.
[0041] In the experimental setup, the SEED dataset was used for a three-class classification task (positive, neutral, negative). The SEED dataset contains EEG records from 15 participants, with 3394 samples for each participant.
[0042] Table 1 Comparison of SEED classification performance
[0043] Table 1 summarizes the average accuracy, standard deviation, F1 score, and Kappa coefficient of each method on the SEED dataset. The F1 score combines the precision and recall of classification, reflecting the model's overall discriminative ability across different emotion categories; the Kappa statistic measures the consistency between the model's predictions and the true labels, further characterizing the reliability and stability of classification performance. On the SEED dataset, the proposed model achieved an average accuracy of 93.71%, an F1 score of 0.94, and a Kappa value of 0.91, significantly outperforming other comparative methods. Compared to the best-performing multi-source alignment model MSSDDA, the proposed method improved recognition accuracy by 3.95%, indicating that the model has stronger performance and generalization ability in capturing the spatiotemporal dependence of EEG signals and cross-subject transfer features. Furthermore, the standard deviation of the proposed method is 5.78, indicating that the model maintains high stability among different subjects.
[0044] Table 2 Ablation experiments on SEED
[0045] To further verify the contribution of each key module in the proposed model to the overall performance, ablation experiments were conducted on the SEED dataset, and the results are shown in Table 2. The experiments progressively removed or replaced core modules from three aspects: graph structure modeling, distribution alignment, and source domain weighting, to evaluate their impact on recognition performance. Overall, the complete model achieved a maximum average accuracy of 93.71%. When the source domain weighting mechanism (PDWM) was removed, the accuracy decreased to 92.16%, a drop of 1.55%, indicating that this mechanism can effectively suppress negative transfer caused by weak source domains and improve the stability of cross-subject emotion recognition. When the conditional distribution weighting module of the branch consistency constraint (BCC) was removed, the model accuracy decreased to 90.95%, a drop of 2.76%, indicating that this module plays a key role in achieving joint alignment of edges and conditional distributions, helping to enhance the discriminative power of cross-domain feature alignment. Furthermore, when the Self-Organizing Sparse Directed Graph Convolutional Network (SS-DirGCN) was replaced with a traditional GCN, the model performance decreased to 92.72%, a drop of 0.99%, which verifies the effectiveness of directed topological modeling in capturing the asymmetric propagation characteristics of EEG signals. Further, when all domain adaptation mechanisms were removed, retaining only the basic feature extraction and classification structure, the model accuracy significantly decreased to 89.59%, a 4.12% decrease compared to the complete model, indicating that the joint distribution alignment and source domain weighting design played a crucial role in cross-subject transfer tasks. In summary, the ablation experiments fully validated the positive contributions of each module to the model performance improvement and demonstrated the rationality and necessity of the proposed model in EEG emotion recognition tasks from different perspectives.
[0046] To further explore the ability of the proposed directed graph convolutional network (DirGCN) in learning EEG spatial connectivity patterns, electrode channel correlation analysis was performed on the brain functional connectivity learned by the model based on the SEED dataset. The directed adjacency matrix obtained from the three-layer directed graph convolution (DirGCN) was averaged to obtain the directed brain functional connectivity map for each subject. Subjects 1 and 6 were selected as typical representatives for visualization and analysis, as shown below. Figure 2 As shown. To highlight key inter-channel interactions, the top ten directed connections with the highest weights for each subject are displayed. For Subject 1, the top ten connections include: C1→PO6, CZ→PO6, C1→F7, CZ→F7, F7→PO6, PO6→F7, PO8→F7, O1→PO6, CB1→PO6, PO8→PO6. Spatially, these connections are mainly concentrated in the occipital and parietal lobe regions, extending to the central and frontal lobe regions. Among them, the bidirectional connection between F7 and PO6 indicates significant fronto-occipital pathway information interaction during emotion induction, suggesting that both the early visual processing area and the higher emotion appraisal area are activated. For Subject 6, the top ten connections include: FC2→F1, FCZ→F1, C3→F7, F2→F1, FZ→F1, AF3→F1, FPZ→F1, OZ→F1, F5→F7, F7→F5. Compared to Subject 1, this subject's connections were primarily concentrated within the frontal lobe region, exhibiting localized and short-range information flow characteristics, with more significant interactions between the prefrontal cortex and the precentral electrode. Specifically, the bidirectional connection between F5 and F7, and the convergent structure with multiple channels pointing to F1, indicate that information is integrated in the frontal pole region, which is considered closely related to individual differences in emotional responses. The results demonstrate that the proposed directed graph convolution framework can effectively capture local and cross-regional brain functional connectivity patterns related to emotional processing. Significant differences in connection distribution among different subjects indicate substantial individual variability in the functional connectivity organization structure under emotional states.
[0047] To further explore the model's attention patterns to different brain regions during emotion feature learning, a significance analysis was performed on the SEED dataset to quantify the importance contribution of each EEG channel in emotion recognition. The channel results for all subjects were averaged and presented as EEG topographic maps, as shown below. Figure 3 As shown in the figure, the overall distribution results indicate that the model shows significantly higher response intensity in the channels of the frontal and temporal lobes, which is highly consistent with the core role of the frontal and temporal lobes in emotion perception and regulation in brain science research. This demonstrates that the model proposed in this invention can not only effectively capture the brain region activity characteristics highly correlated with emotional states, but also present spatial activation patterns consistent with psychological and neuroscience research across the entire brain.
Claims
1. A multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution, characterized in that, Includes the following steps: S1. Acquire and preprocess EEG signals; Constructing brainwave feature vectors based on brainwave signals; S2. Construct a sparse adjacency matrix based on EEG feature tensors; S3. Perform directed graph convolution operations on the forward and reverse adjacency matrices of the sparse adjacency matrix respectively to capture the asymmetric information propagation characteristics between EEG channels and generate domain-invariant EEG emotional features. S4. Construct multiple domain feature extraction branches for multiple source domain samples and target domain samples; The domain-invariant EEG emotion features are input into multiple domain feature extraction branches respectively; Construct an independent feature representation space for each source and target domain, separating domain-shared features from domain-specific features; The classifier classifies samples based on domain-shared features and domain-specific features; S5. Integrate the prediction results of each source domain classifier for the target domain samples and output the EEG emotion recognition result of the target domain; use the weighted result of the overall loss function of each source domain to perform cross-domain optimization of the model.
2. The multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution as described in claim 1, characterized in that, S1 includes: The acquired raw EEG signals were bandpass filtered and divided into frequency bands according to different datasets. The signal variance of each EEG channel in each frequency band was calculated within a non-overlapping time window, and the differential entropy features of the EEG signals were extracted based on the differential entropy theory. The differential entropy features were then subjected to time smoothing using a linear dynamic system to obtain the EEG feature tensor.
3. The multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution according to claim 1, characterized in that, S2 include: S21. Map the EEG feature tensor to the latent space to generate the mapped feature matrix; S22. Construct an adjacency matrix based on the mapped feature matrix to represent the adjacency relationship between channels; S23. Construct a sparse adjacency matrix based on the adjacency matrix.
4. The multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution as described in claim 3, characterized in that, S23 includes: A Top-k sparsity strategy is applied to the adjacency matrix, retaining only the channel with the largest connection weight. Connect 1, and reset the weights of the remaining connections to zero, constructing a sparse adjacency matrix: in, This represents Hadamard element-wise multiplication. Let A represent the sparse mask matrix, and A be the adjacency matrix.
5. The multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution according to claim 1, characterized in that, S3 includes: S31. Based on the sparse adjacency matrix, define the forward adjacency matrix and the reverse adjacency matrix, and then construct the direction-sensitive message passing matrix. S32. Feature extraction of the message passing matrix is performed at multiple scales through forward propagation, backward propagation, and self-loop information. Finally, the multi-scale features are fused through an attention mechanism to obtain domain-invariant EEG emotion features.
6. The multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution according to claim 5, characterized in that, In S32, the feature update process is represented as follows: in, It is the first Layer output; , , These are the forward weight matrix, the reverse weight matrix, and the self-loop weight matrix, respectively. It is the fusion coefficient, used to balance the forward and reverse information flows; It represents the characteristic information propagating from the source node to the target node, reflecting the strength of the forward flow of the signal; It indicates the reverse flow of information.
7. The multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution according to claim 1, characterized in that, S5 include: S51. Construct the overall loss function; S52. Based on the prediction results of each source domain classifier for the target domain samples, the prediction confidence is obtained: the average of each prediction confidence is taken as the source domain transfer performance index; the source domain weight coefficient is obtained by using the exponential amplification function based on the source domain transfer performance index. S53. The overall loss function of each source domain branch is weighted according to the source domain weight coefficient to achieve cross-domain optimization; the prediction results of each source domain are fused by weighting the source domain weight coefficient to output the EEG emotion recognition result of the target domain.
8. The multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution according to claim 7, characterized in that, S51 includes: Construct a multi-branch consistency constraint loss to constrain the prediction consistency of the target domain across different source domain branches: in, The number of source domain classifiers, For the first The predicted vectors of each branch in the target domain.
9. The multi-source domain EEG emotion recognition method based on self-organizing sparse directed graph convolution according to claim 8, characterized in that, In S5, the overall loss function is: in, , , For loss weight hyperparameters, For source domain supervised classification loss, The edge distribution alignment loss is based on the normalized Wasserstein distance. This is the conditional distribution alignment loss based on the local maximum mean difference. The loss is due to the multi-branch consistency constraint.