The Research On Imbalanced Data Classification In Brain-Computer Interface

Brain-computer interface(BCI)technology is an assistive technology that conveys users’ intentions by decoding various brain activities and translating them into control commands,essentially is the use of thinking control external devices human-computer interaction technology.The recognition error is inevitable in BCI systems,which will affect their performance greatly,then in turn limit the development and application of BCI technology.Therefore,it’s necessary to detect the Error Related Negativity(ERN)in BCI systems.Because of the non-equilibrium of ERN datasets,traditional machine learning algorithms are skewed to most classes and the samples of the few classes are misclassified.Although the overall classification accuracy is relatively high,the classification accuracy of the minority class is very low and the classification efficiency is not ideal.To solve this problem,this paper studies the two aspects of data processing and classification algorithm,the main work is as follows:Because the ERN only exists in the error monitoring process and has little relationships with specific paradigms,it can be integrated into the BCI system in other modes,recording the EEG for the second time and recognizing the existence of ERN,so as to correct the mistakes and improve the BCI’s reliability.Because the single trial detection of ERN is difficult and multichannels data may cause over fitting,the paper proposed a new approach of extracting the EEG temporal features,and to overcome the shortage of high dimension of features,the local linear embedding(LLE)is used to reduce the dimensionality of multi-channels and enhance the expression ability of features.Although the ERN signal is inter-individual different,the classification results using receiver operating characteristic curves(ROC)and area under ROC curve(AUC)metrics suggest the classification accuracy of the minority class increased by nearly 4 times.(AUC: 0.7566).At present,the best way to solve the problem of imbalanced data classification is from two aspects: data and algorithm.it is helpful to improve the classification prediction rate of imbalanced data by selecting valuable features and selecting a classifier suitable for data feature.For the traditional support vector machine(SVM),when dealing with imbalanced datasets with large sample imbalance,the support vectors of a few classes are neglected,and the decision boundaries are enlarged,resulting in the final decision surface deviated and the classification results(ERM)based on the Extreme Learning Machine(ELM),this paper designs an ERN detection system based on Extreme Learning Machine(ELM).Firstly,deduce and discuss the principle of the algorithm.Then,the simulation results are analysed respectively from training time,classification accuracy,ROC and AUC.The experimental results show that the total classification accuracy of the ELM classifier is 90.72% compared with the traditional SVM classifier,while the average AUC is as high as 0.8377,and the training speed is improved by nearly 10 times,showing better classification performance.