A Fast, Efficient Domain Adaptation Technique for Cross-Domain Electroencephalography(EEG)-Based Emotion Recognition

Abstract

Electroencephalography (EEG)-based emotion recognition is an important element in psychiatric health diagnosis for patients. However, the underlying EEG sensor signals are always non-stationary if they are sampled from different experimental sessions or subjects. This results in the deterioration of the classification performance. Domain adaptation methods offer an effective way to reduce the discrepancy of marginal distribution. However, for EEG sensor signals, both marginal and conditional distributions may be mismatched. In addition, the existing domain adaptation strategies always require a high level of additional computation. To address this problem, a novel strategy named adaptive subspace feature matching (ASFM) is proposed in this paper in order to integrate both the marginal and conditional distributions within a unified framework (without any labeled samples from target subjects). Specifically, we develop a linear transformation function which matches the marginal distributions of the source and target subspaces without a regularization term. This significantly decreases the time complexity of our domain adaptation procedure. As a result, both marginal and conditional distribution discrepancies between the source domain and unlabeled target domain can be reduced, and logistic regression (LR) can be applied to the new source domain in order to train a classifier for use in the target domain, since the aligned source domain follows a distribution which is similar to that of the target domain. We compare our ASFM method with six typical approaches using a public EEG dataset with three affective states: positive, neutral, and negative. Both offline and online evaluations were performed. The subject-to-subject offline experimental results demonstrate that our component achieves a mean accuracy and standard deviation of 80.46% and 6.84%, respectively, as compared with a state-of-the-art method, the subspace alignment auto-encoder (SAAE), which achieves values of 77.88% and 7.33% on average, respectively. For the online analysis, the average classification accuracy and standard deviation of ASFM in the subject-to-subject evaluation for all the 15 subjects in a dataset was 75.11% and 7.65%, respectively, gaining a significant performance improvement compared to the best baseline LR which achieves 56.38% and 7.48%, respectively. The experimental results confirm the effectiveness of the proposed method relative to state-of-the-art methods. Moreover, computational efficiency of the proposed ASFM method is much better than standard domain adaptation; if the numbers of training samples and test samples are controlled within certain range, it is suitable for real-time classification. It can be concluded that ASFM is a useful and effective tool for decreasing domain discrepancy and reducing performance degradation across subjects and sessions in the field of EEG-based emotion recognition.

A Review of Generalizable Transfer Learning in Automatic Emotion Recognition

“…Commonly used transfer learning applications attempt to train subject-specific (personalized) models by providing knowledge learned from subjects similar to the one in the test (Lin and Jung, 2017;Lin, 2019), or simply consider all the members in the group by assigning different weights . Other methods include statistical approaches, such as Principal Component Analysis (PCA) and adaptive subspace feature matching (ASFM) (Chai et al, 2017). As the development of wearable devices progresses and more physiological data related to emotional experiences become available, transfer learning methods appear to depict a great potential in this domain.…”

Mental Workload Estimation Using Wireless EEG Signals

“…ASFM was proposed in [28] as a fast domain adaptation technique for EEG-based emotion recognition to overcome the degradation of algorithm when EEG data are sampled from different subjects or sessions. The nonstationary nature of EEG and variability of brain dynamics with individuals and age causes a mismatch between the marginal and conditional distributions of the source domain (training data) and target domain (testing data).…”

“…In other words, if there is a source domain Xs with label Ys and a target domain Xt with label Yt, ASFM formulates a new feature to reduce the marginal distribution mismatch between Ps(Xs) and Pt(Xt), and conditional distribution mismatch between Ps(Ys|Xs) and Pt(Yt|Xt). For details of the algorithm, see [28].…”

“…This result suggests that the model with ASFM could be used for a subject at every session without retraining. ASFM was first used in [28] for emotion recognition using differential entropy as features. In that work, it achieved a cross-session accuracy of 75.1% (SD = 7.7%).…”

“…The model was tested across eight (8) subjects in two different types of taskn-back task and arithmetic task. In addition, a fast domain adaptation technique called Adaptive Subspace Feature Machine (ASFM) [28] was applied to improve the model performance in cross-session, cross-task and crosssubject classifications. We compared the results from ASFM with those of SVM.…”