Effects of feature reduction on emotion recognition using EEG signals and machine learning

Abstract

AbstractElectroencephalography is a core technology of brain computer interfaces. Even a few number of electrodes can produce complex signals that are difficult to interpret. This is particularly true when trying to detect complex mental states, such as the identification of human emotions. This work analyzes the impact of feature reduction using the SJTU Emotion EEG data set, considering inter‐subject and inter‐session EEG classification. It is well established that EEG data tends to be subject dependent, and that the brain patterns produced by a person in response to a specific stimuli will tend to change over their lifetime, we can say that the above classification task is one of the most challenging test scenarios. We extract 1068 time and frequency domain features, single and multiple channels. Feature reduction was analyzed using both filter and wrapper based approaches. In the first group we use principal component analysis, Kernel PCA and Locally Linear Embedding, and in the second three meta‐heuristic algorithms are employed. For classification, we apply multi‐layer perceptron, quadratic discriminant analysis, random forest and support vector machines. Results show up to 100% accuracy on the median test error for some splits in the 10‐fold cross‐validation process. Best results were achieved by filter‐based selection, with the best average accuracy reached by Random Forest (93.20%) and Multi‐Layer Perceptron (92.71%), using components found with Kernel PCA. To the best of our knowledge, these classification results are the best ever achieved for the SEED data set on the subject‐independent and session‐independent problem. Comparisons using a Leave‐One‐Subject‐Out cross‐validation setting shows that the proposed methodology achieves state‐of‐the‐art results using only 12 electrodes. These off‐the‐shelf ML techniques, coupled with filter based feature reduction, achieve the best results, outperforming complex classification models and wrapper based meta‐heuristics.