Compact unsupervised EEG response representation for emotion recognition

In this work, we propose a compact and un-supervised EEG response representation. Instead of directly extracting features from the whole response, as is commonly done for EEG signal processing, the proposed representation employs segment-level feature extraction and leverages a robust two-part unsupervised generative model to transform the segment-level features to a low-dimensional vector. The proposed method leads to rich and compact representation capability, and robust unsupervised estimation. While some previous work [1] based on segment-level features needs labeled training responses to transforms segment-level features to a response representation, the proposed method produces an EEG response representation in an unsupervised fashion, which can be directly used in various EEG response classification problems. We perform binary classification and regression of emotion dimensions on the DEAP dataset (Dataset for Emotion Analysis using electroencephalogram, Physiological and Video Signals) and demonstrate competitive performances.