TL;DR
This paper introduces GASDSL, a novel semi-supervised learning framework for EEG-based emotion recognition. GASDSL jointly learns a discriminative subspace and an adaptive graph to improve accuracy by simultaneously reducing dimensionality, enhancing feature separability, and leveraging unlabeled data.
What Problem Does This Paper Address?
Recognizing human emotions from Electroencephalogram (EEG) signals is a challenging but crucial task for developing affective brain-computer interfaces. While EEG offers rich neurophysiological information, it's inherently non-stationary and prone to noise, making analysis difficult. Existing methods often struggle with high-dimensional EEG data, which can contain redundant and noisy features. Furthermore, many approaches treat emotion recognition as a pure pattern classification task, neglecting the underlying neural mechanisms. This paper aims to address these limitations by proposing a method that not only improves recognition accuracy but also provides insights into the critical EEG features and brain regions involved in emotion expression.
Methods
The core of the proposed method is the Graph Adaptive Semi-supervised Discriminative Subspace Learning (GASDSL) model. GASDSL unifies three key components into a single optimization objective:
- Discriminative Subspace Learning: EEG data is projected into a lower-dimensional subspace to reduce dimensionality and enhance the separability between emotional states. This is achieved by minimizing the within-class scatter while maximizing the between-class scatter.
- Adaptive Maximum Entropy Graph Learning: A graph similarity matrix is constructed to capture the local structural information of EEG data. An adaptive maximum entropy regularizer is employed to prevent trivial solutions and ensure that the graph accurately reflects sample similarities, especially within-class connections.
- Semi-supervised Emotional State Prediction: The model leverages both labeled and unlabeled EEG data. It iteratively updates pseudo-labels for unlabeled samples and estimates emotional state centroids, effectively learning from the available data.
These three components are jointly optimized, allowing them to influence and improve each other. The model iteratively updates the projection matrix (W), centroid matrix (M), label indicator matrix (F), and graph similarity matrix (S) until convergence.
Datasets
The proposed GASDSL model was evaluated on two publicly available benchmark emotional EEG datasets:
- SEED-IV: This dataset contains EEG recordings from 15 healthy subjects elicited by four emotional states (sad, fear, happy, neutral).
- SEED-V: This dataset includes EEG data from 16 subjects experiencing five emotional states (fear, sad, neutral, happy, disgust).
Both datasets utilize Differential Entropy (DE) features extracted from five frequency bands (Delta, Theta, Alpha, Beta, Gamma) across 62 EEG channels, resulting in a feature dimensionality of 310. The experiments were conducted using a cross-session evaluation strategy, where data from one session was used for training (labeled) and data from another session was used for testing (unlabeled).
Key Insights and Ideas
- Unified Optimization: The key innovation is the joint optimization of discriminative subspace learning, adaptive graph construction, and semi-supervised label prediction. This holistic approach allows each component to benefit from the others, leading to improved performance.
- Adaptive Graph for Structure: The use of a maximum entropy regularizer for graph construction is crucial. It adaptively determines neighborhood sizes, making the graph more robust to noise and better at capturing local data structure, which is vital for EEG analysis.
- Beyond Classification: GASDSL not only aims for high recognition accuracy but also provides a mechanism to identify important EEG features and brain regions. This offers valuable insights into the neural correlates of emotion.
- Out-of-Sample Capability: Unlike transductive methods, GASDSL's learned projection matrix allows for direct prediction on new, unseen EEG samples, making it more practical for real-world applications.
Key Results
- Superior Accuracy: GASDSL consistently achieved the highest average emotion recognition accuracies across all cross-session tasks on both SEED-IV and SEED-V datasets, outperforming several state-of-the-art semi-supervised learning models. For instance, on SEED-IV, GASDSL improved performance by up to 4.59% compared to the second-best model.
- Discriminative Subspace: Visualizations on synthetic and real EEG data demonstrated that GASDSL effectively learns a discriminative subspace where different emotional states are clearly separated, even in the presence of significant noise.
- Feature Importance: Analysis of the learned projection matrix revealed that features from the Gamma band and specific brain regions including the left/right temporal, prefrontal, and (central) parietal lobes were most critical for emotion recognition. This aligns with existing neuroscientific findings.
- Parameter Sensitivity: The model showed some sensitivity to its parameters, particularly $\beta$ (related to entropy regularization), suggesting that careful tuning is important for optimal performance.
Limitations and Future Work
- Computational Complexity: While GASDSL offers improved performance, its computational complexity is noted as $O(tn^2d)$, where $t$ is the number of iterations, $n$ is the number of samples, and $d$ is the feature dimensionality. This could be a consideration for very large datasets.
- Cross-Subject Variability: The current work focuses on cross-session recognition. The authors suggest future work on addressing inter-subject variability in cross-subject EEG emotion recognition, potentially using bipartite graph-based adaptive data transferability.
- Multimodal Data: Extending GASDSL to handle multimodal data (e.g., combining EEG with facial expressions or speech) is another promising avenue for future research.
Summary updated