Empowering computer systems to automatically recognize human emotions has become an urgent need in the field of human-computer interaction (HCI). Two-dimensional emotion (Valence-Arousal) models are commonly used to represent emotions. Up to now, the correlation between emotion dimensions has rarely been investigated, and subject-independent EEG emotion recognition is still a challenging task. For this purpose, we introduce multi-task learning (MTL) into EEG emotion recognition. MTL learns different emotion dimensions simultaneously and extracts correlation information between dimensions in task-sharing space to coordinate the optimization of multiple emotion dimensions. We further propose Identity based Multi-gate Mixture-of-Experts (IDMMOE), which allocates part of model subspace for each subject in a customized manner according to the subject’s identity. Extensive experiments were conducted on DEAP dataset. Three MTL models were implemented: Shared-Bottom, Multi-gate Mixture-of-Experts, and Customized Gate Control respectively. They were compared with a single-task learning model trained separately on valence and arousal. Experimental results demonstrate that two emotion dimensions are intrinsically related, and MTL acquires such correlation information and improves prediction accuracy in both emotion dimensions. In addition, IDMMOE achieves average accuracies of 89.5% and 89.7% for valence and arousal respectively and it is effective for subject-independent experiment.

Electroencephalography (EEG)-based emotion recognition methods mostly adopt features as many as possible to achieve high performance. However, this is not feasible and flexible for portable or wearable helmet devices because of the limited computing resource and requirement of real-time performance. Aiming at improving EEG emotion recognition performance with features as few as possible, this paper firstly designed a feature called Relative Intensity Ratio Entropy (RIRE), then proposed a novel feature selection method based on greedy algorithm and Max-Relevance and Min-Redundancy(mRMR), termed as Greedy-mRMR. Greedy algorithm is adopted to take top features in the ranking list, and mRMR algorithm is used to select features that are most relevant and minimum redundant. Greedy-mRMR algorithm also uses dynamic termination threshold to guarantee recognition accuracy in each iteration. Experiments were carried on DEAP dataset. SVM classifier achieved average classification accuracy of 91.16% in valence and 91.70% in arousal, by extracting RIRE feature and selecting with Greedy-mRMR algorithm. Experimental results show that RIRE is suitable for EEG emotion recognition and Greedy-mRMR outperforms state-of-the-art methods.

Emotion classification plays a critical role in the development of human-computer interaction. EEG (Electroencephalogram) signal is an important information source, from which different features have been extracted for the study of emotion states. However, there is a large amount of redundant and irrelevant information in EEG signals, which directly interferes with emotion classification. Aiming at selecting EEG emotional features precisely and efficiently, this paper proposed a feature selection framework, termed MIGA (Mutual Information and Genetic Algorithm). It combined mutual information and genetic algorithm to improve the quality of feature subset based on three perspectives, that is correlation, contribution and synergistic effect between features. A new emotional feature IQI (Intensity Quantity of Information) was also designed in this paper. IQI is able to mine the intensity information of EEG signals, and enlarge the samples’ distance as a result. Experiments were carried on DEAP (A Database for Emotion Analysis using Physiological Signals). Results show that the classification accuracy of IQI is 5% higher than that of the traditional frequency domain feature, and MIGA reduces feature amount by 2/3 while ensuring classification accuracy. For DEAP, classification accuracy of MIGA in valence and arousal reached 88.55% and 88.14% respectively. It is indicated that, compared with existing methods, MIGA improves emotion recognition with much fewer features from EEG.

EEG signals directly reflect the active state of the brain, so they are widely used for emotion recognition. At present, many researchers have achieved noteworthy results by using multi-channel EEG signals. However, too many EEG channels will cause slow transmission, high experimental costs, and low efficiency. This paper proposed a grouped dynamic EEG channel selection method based on ReliefF and random forest (RF), termed GDCSBR. We divided all channels into four groups, which provided more choices and flexibility to subsequent dynamic channel selection. GDCSBR selected channels iteratively. With the iteration increased, the number of alternative channels decreased. At each iteration, we adopted the strategy with a minor loss of recognition accuracy. Finally, the results on subject-independent data were taken as the final choice, since there were relatively large differences between subjects. Experiments were carried out on the DEAP dataset. The recognition accuracy for valence reaches 81.27% while 10 channels are selected. As for arousal scale, 11 channels can obtain 82.36% of classification accuracy. In addition, we found that high-frequency bands play a crucial part in emotion recognition, and the selected channels were mostly located in the frontal and parietal lobes. These findings are coincident with previous work. Experimental results demonstrate the effectiveness of the proposed method.