Scientists Develop Neural Network Methods to Analyze Emotions in Neurophysiological Data

Researchers at Moscow State University (MSU) have introduced a new family of machine learning model architectures using graph neural networks. This innovation promises to significantly improve the interpretation and generalization capabilities when analyzing complex, multidimensional time series data, as well as to incorporate additional metric information from specific subject areas. The results of the study were presented at the All-Russian Conference on “Mathematical Methods of Pattern Recognition.”

Faculty members from the MSU Faculty of Computational Mathematics and Cybernetics formalized the tasks of identifying functional patterns in multidimensional time series using machine learning methods. This approach eliminates the need for specialized research in each specific field. The proposed architecture takes spatial metadata into account by encoding the original positions of electrodes as a graph, which is then fed into the corresponding model. This method was successfully applied in neurophysiology: the model independently handled cases where neurophysiologists already recognized the functional P300 pattern. Additionally, this approach was used for emotion recognition with the SEED dataset, achieving high results.

“One of the key aspects of the method is the use of a graph constructed from external metric information, which had a significant impact on the algorithm’s effectiveness. In the future, we plan to expand this method to other areas, such as data from sensors on production lines or banking transactions,” noted Archil Maisuradze, Associate Professor at the Department of Mathematical Methods of Forecasting, MSU.

The research covers the analysis of datasets collected using the “single stimulus” experimental method, which is widely used in psychological and neurophysiological studies. The BCI competition dataset was used for additional testing of the architecture, achieving impressive results with various machine learning models. Furthermore, the SEED dataset was used for emotion recognition tasks, demonstrating the flexibility and effectiveness of the method.

“Unlike recent approaches that ignore the physical structure of BCI interfaces, this method emphasizes creating a dense graph representing the actual shape of the recording device. We explored different methods of graph construction and showed that our proposed method significantly affects the results,” added Leonid Sidorov, a graduate student at the Department of Mathematical Methods of Forecasting, MSU.

The proposed model architecture consists of spatial and temporal processing blocks, as well as a prediction block, with a focus on the capabilities of graph convolutional networks. Integrating graph convolutional networks led to improved model performance on EEG data and competitive accuracy levels on standard datasets.