TLDR: ADSEL (Adaptive Dual Self-Expression Learning) is a new algorithm for EEG-based emotion recognition that addresses the challenge of incomplete multi-dimensional emotional labels. It improves feature selection by integrating information from both data samples and emotional dimensions to accurately reconstruct missing labels. Tested on public datasets, ADSEL demonstrated superior performance compared to existing methods, particularly when dealing with partial label absence, offering a more robust solution for real-world applications.
Electroencephalogram (EEG) signals, which capture brain activity, are increasingly vital for understanding emotions in human-computer interactions. However, a significant challenge in this field is the high volume of EEG data, often combined with limited sample sizes. This can lead to models that are too specialized for the training data (overfitting) and require extensive computational power. Feature selection, the process of choosing the most relevant parts of the data, is a crucial step to overcome these hurdles.
Many existing methods for selecting EEG features assume that all emotional labels for the data are complete. In reality, collecting EEG data in open environments and the subjective nature of emotional experiences often result in incomplete or missing labels. This incompleteness can severely impact how well a model generalizes to new data. Furthermore, current feature selection techniques that handle incomplete labels tend to focus only on the relationships between different emotional dimensions when trying to fill in missing information. They often overlook the connections between individual data samples and how these samples interact with various emotional dimensions.
Introducing ADSEL: A Novel Approach
To address these limitations, researchers have proposed a new algorithm called Adaptive Dual Self-Expression Learning (ADSEL). This innovative method is specifically designed for EEG-based emotion recognition when dealing with incomplete multi-dimensional emotional labels. ADSEL integrates an adaptive dual self-expression learning process with least squares regression, creating a unique bidirectional pathway.
This bidirectional pathway allows for a dynamic exchange of learned information between how individual data samples relate to each other and how different emotional dimensions relate to each other within the label space. By doing so, ADSEL can simultaneously leverage effective information from both samples and dimensions to reconstruct missing labels more accurately. This enhanced label recovery, in turn, helps in identifying the most optimal subset of EEG features for multi-dimensional emotion recognition.
How ADSEL Works
The ADSEL framework is built upon three core components: basic feature learning, adaptive dual self-expression learning, and global feature redundancy learning. Basic feature learning establishes the relationship between the EEG data and the reconstructed labels. The adaptive dual self-expression learning module is the heart of ADSEL, enabling the recovery of missing labels by integrating information across both samples and dimensions. This is crucial for handling the real-world challenge of incomplete data. Finally, global feature redundancy learning helps to minimize redundant information among the selected EEG features, ensuring that the chosen subset is both informative and efficient.
Rigorous Evaluation and Promising Results
The effectiveness of ADSEL was rigorously tested against 14 other state-of-the-art feature selection algorithms. The evaluation was conducted on two widely recognized public EEG datasets, DREAMER and DEAP, both of which contain multi-dimensional emotional labels. These datasets allowed for a comprehensive assessment of ADSEL’s performance under various conditions of partial label absence, ranging from 10% to 50% missing labels.
The experimental results clearly demonstrated ADSEL’s superior performance, especially when dealing with partially missing labels. Across various evaluation metrics, ADSEL consistently achieved better emotion recognition performance compared to the other benchmarked methods. This indicates that the EEG feature subsets selected by ADSEL are highly effective for incomplete multi-dimensional emotion recognition tasks. Furthermore, the research showed that ADSEL is robust and not overly sensitive to changes in its internal parameters, and its optimization algorithm converges quickly, ensuring practical applicability.
Also Read:
- FDC-Net: A Collaborative Approach to EEG Emotion Recognition in Noisy Environments
- IMAC: Enhancing EEG Signal Classification Through Spatial Imputation
Looking Ahead
While ADSEL marks a significant advancement, the researchers acknowledge that its label recovery performance can still be constrained by very high label missing rates and extremely limited sample sizes. In such challenging scenarios, capturing the complex relationships between samples, dimensions, and their interactions within the label space remains difficult. Future work will focus on improving label recovery accuracy under these more extreme conditions, further enhancing the capabilities of EEG-based emotion recognition systems. You can read the full research paper here: ADSEL: Adaptive dual self-expression learning for EEG feature selection via incomplete multi-dimensional emotional tagging.
