Abstract
Electroencephalogram (EEG) signals are characterized by significant inter-subject and intra-subject variability, posing a major challenge for Brain-Computer Interface (BCI) systems. This paper introduces Group Nonnegative Matrix Factorization (GNMF), a novel approach that simultaneously analyzes EEG data from multiple subjects. GNMF decomposes signals into "common" bases representing shared task-related patterns and "individual" bases capturing subject-specific noise and anatomy. Experimental results on real-world BCI datasets demonstrate that GNMF significantly outperforms standard NMF variants and competition winners, particularly in subject-to-subject transfer learning scenarios.
Introduction
The primary hurdle in developing robust BCIs is the inherent "noise" and inconsistency of EEG data. While specific mental tasks, such as motor imagery, elicit well-known rhythms (e.g., $\mu$ and $\beta$ oscillations), their exact spectral and spatial signatures vary across individuals and even across different sessions for the same individual. Standard techniques like Nonnegative Matrix Factorization (NMF) often fail to generalize because they either treat subjects in isolation—ignoring group-level insights—or aggregate them naively, which obscures individual nuances.
Methodology: The GNMF Framework
The authors propose GNMF to bridge the gap between individual and group analysis. In a standard NMF setup, a data matrix $X$ is decomposed into a basis matrix $A$ (features) and an encoding matrix $S$ (activations). GNMF extends this by partitioning the basis matrix $A$ for each subject $l$ into two distinct sub-matrices:
- Common Bases ($A_C$): Features expected to be similar across the group, representing the underlying neural correlates of the task.
- Individual Bases ($A_I$): Features that capture unique physiological traits or localized artifacts.
To enforce this separation, the model utilizes a specialized objective function with two key regularizers:
- Similarity Constraint ($\alpha$): Minimizes the distance between the common bases of different subjects.
- Dissimilarity Constraint ($\beta$): Maximizes the distance between individual bases, ensuring that unique subject traits are not erroneously grouped into the common category.
Visualizing Commonality and Individuality
The effectiveness of the GNMF approach is best illustrated by examining the extracted components across multiple subjects. As shown in the figure below, the "Common Basis Components" exhibit high temporal correlation across subjects, effectively capturing the shared "essence" of the mental task. Conversely, the "Individual Basis Components" show distinct, uncorrelated patterns that represent subject-specific noise or anatomical variations.
Key Results and Performance
The GNMF method was evaluated using the IDIAP dataset from BCI Competition III. Key findings include:
- Superior Classification: Using only the common features, GNMF achieved an average accuracy of 74.51%, surpassing the competition winner (68.65%) and fixed-basis models like FFX-NMF (70.34%).
- Enhanced Transfer Learning: GNMF proved exceptionally robust for subject-to-subject transfer. When trained on two subjects and tested on a third, GNMF maintained high accuracy (e.g., 69.12% for Subject 3), whereas naive concatenation methods dropped to near-chance levels (47.71%).
- Noise Filtering: By explicitly modeling individual variation, the common bases act as a "social" regularizer, allowing the model to learn cleaner task representations by observing patterns across the entire group.
Conclusion
Group Nonnegative Matrix Factorization provides a powerful mathematical framework for multi-subject EEG analysis. By balancing shared group wisdom with individual flexibility, it addresses the core challenge of inter-subject variability. Future work may involve automating hyperparameter selection through Bayesian methods and exploring the scalability of the algorithm for larger clinical or consumer-grade EEG datasets.
Summary updated