Two-interaction iterative multi-layer classification model for EEG signals using support vector machines

The classification of Epileptic Electroencephalogram (EEG) signals by machine learning has become one of the current research hospitals. The research work can be roughly divided into two stages. (1) How to extract effective training features from the original signal; (2) How to construct or train the appropriate model according to the existing training features. However, it is not easy to establish such an appropriate training model. In this study, we propose a two-interactive iterative multi-layer modeling learning method based on classical support vector machine (SVM). In order not to excessively increase the extra computational cost, we set two SVMs in a training-module for parallel calculation and mutual supervision and adjustment. The training stop conditions are set, and the outputs of two SVMs are used to determine the number of model iterative training, which gives full play to the classification advantages of each SVM and alleviates the overfitting problem. A training sample space optimization method is proposed, which considers the mutual guiding effect of decision-making information between different training-modules and different SVMs in the same module, and realizes the consistency of the model with progressive training mode. In the end, the proposed model wins the second place in most of the constructed datasets, with its best training accuracy of 97.11% and the best testing accuracy of 96.06%, which also confirms the feasibility of the proposed model.