The Study On EEG-Based Emotion Classification Based On Symbolic Representation Learning

Electroencephalogram(EEG) records the process of electric potentials corresponding to brain activities. It reflects the cranial nerve physiological activities in the cerebral cortex. Thus, it plays an important role in many complicate applications such as emotion computing and brain cognitive mechanism research. Traditional EEG analysis methods majorly utilize the global features from time-domain, frequency-domain and time-frequency-domain, respectively. However, these methods are hard to reflect the local and time segment characteristics of EEG sequences. In the recent years, EEG feature extraction methods based on information theory, fractural dimension and complexity analysis have been developed for different EEG-related analysis tasks for their good stability.This study investigates the techniques for extracting effective EEG features from EEG-based emtoion classification. The main works including: Firstly, several baseline EEG-based emotion classification systems are developed based on some typical features including entropy, fractal dimension, and complexity based features. Considering that the performances of these individual baseline methods are not satisfactory, the multiple feature fusion strategy is expected to further improve their performances. It is observed that the direct multiple feature fusion method is hard to improve the performance since this method loss the non-linear structure between these features. Thus, in this study, we propose a decision fusion method based on multiple features in EEG-based emotion classification task. The confidence of each baseline classifier on each class is estimated. These confidences are utilized to integrate the outputs of base classifiers by following a weighted fusion strategy. The experiments on DEAP dataset(EEG emotion classification) show that this method improves the Valence classification for 1.5% and the Arousal classification for 0.8%. Secondly, traditional EEG-based emotion classification methods quantitative EEG signals as one or multiple real value features based on global analysis. This leads the loss of many local characteristics of EEG signals. Thus, traditional methods have performance bottleneck. Target to this problem, in this study, we propose a EEG-based emotion classification method based on symbolic representation learning of EEG signals. Through the discrete process on original EEG signals, these signals are transferred to a new feature space, namely, character strings, through symbolic representation of EEG. On this basis, the N-gram model, topic discovery and anomaly detection methods are applied to extract effective feature related to different emotions from the EEG character strings. Finally, the bag of words(BOW) representations is adopted to classify the emotions corresponding to original EEG signals. The experiments on DEAP dataset show that this method achieves the state of art performance which improves the accuracy on Valence classification for nearly 9%. Meanwhile, the achieved accuracy on Arousal classification is similar to the decision fusion based method.