Cross-subject EEG emotion classification based on few-label adversarial domain adaption

• Published in: Expert systems with applications Vol. 185; p. 115581
• Main Authors: Wang, Yingdong, Liu, Jiatong, Ruan, Qunsheng, Wang, Shuocheng, Wang, Chen
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 15.12.2021; Elsevier BV
• Subjects: Adaptation; Big Data; Cross-subject; Datasets; Deep learning; Distillation; Domains; Electroencephalogram (EEG); Electroencephalography; Emotion classification; Emotions; Few label adversarial domain adaption; Indexing; Machine learning; Signal classification; Electroencephalogram (EEG); Cross-subject; Emotion classification; Few label adversarial domain adaption
• DOI: 10.1016/j.eswa.2021.115581
• ISSN: 0957-4174

Emotion classification signal based on the electroencephalogram (EEG) is an important part of big data associated with health. One of the main challenges in this regard is the varying patterns of EEG indifferent subjects. Domain adaptation is an effective method to reduce the data difference between the source domain and the target domain. However, it is an enormous challenge to make a discriminator-based domain adaptation with a small target data and transform the target domain to the source domain. In the present study, a novel method called “few-label adversarial domain adaption” (FLADA) is proposed for cross-subject emotion classification tasks with small EEG data. The proposed method involves three steps: (a) Selecting subjects of the close source domain forming an adapted list. Few labeled target data are tested based on each emotion model of the source subject to get the subject list of the source domain. (b)Training three models based on each selected subject and the target subject. Three loss functions and six groups’ dataset are designed to get a domain adaption model for each selected source subject. (c) Distilling all classifiers for classifying the target emotion. In general, the main purpose of the proposed method, which originates from the Meta-learning, is to find a feature representation that is broadly suitable for the target subject and source subject with limited labels. The proposed method can be applied to all deep learning oriented models. In order to evaluate the performance of the proposed method, extensive experiments are carried out on SEED and DEAP datasets, which are public datasets. It is found that with a small amount of target data, the proposed FLADA model outperforms the state-of-art methods in terms of accuracy and AUC-ROC. All codes generated in this article are available at github: https://github.com/heibaipei/FLADA. [Display omitted] •Groups forming between source and target data tackles the small data adaption.•A shared feature extractor is proposed between the target model and the source model.•Multi-source domain adaption obtains the best results with a proper number of source.•Six groups with six labels maintains the high accuracy.