Formative semi-supervised learning based on adaptive combined model for brain–computer interface

The recognition of Electroencephalogram (EEG) signals has been an important research field in Brain–computer interface. The semi-supervised classification can improve the classification performance of EEG. Formative Semi-Supervised Learning (FSSL) uses the affinity matrix between samples and Expectation-maximization (EM) to mine hidden features between samples. It isn’t effective to apply FSSL to EEG classification directly due to the non-stationary and nonlinear of EEG. FSSL only uses Euclidean distances in the affinity matrix, which is not sufficient to process EEG signals and may restrict the effect of subsequent feature extraction. In response to this problem, combined model formative Semi-Supervised Learning (CMFSSL) was proposed to construct a combined model based on Euclidean metric and Riemannian metric. The weight update strategy is designed to constrain the model in the EM algorithm, and the weights of the combined model are constantly adjusted to construct a better basic model. Then the hidden features extracted based on the combined model are used to construct the training set and the Broad Learning System is used for classification. The algorithm is verified on three BCI data sets and compared with several state-of-the-art methods. The experimental results show that the algorithm achieves better results on three data sets: 74.86%, 73.52%, 75.49% and has a good effect on cross-domain classification. The combined model uses adaptive weights to build a better data model for subsequent hidden features, which not only maintains the original security advantages, but also improves the classification results.