MMPosE: Movie-Induced Multi-Label Positive Emotion Classification Through EEG Signals

Abstract

2. M M P o s E: M ovi e-in d u c e d m ul tila b el p o si tiv e e m o tio n cl a s sific a tio n t h r o u g h E EG si g n al s. IE E E Tr a n s a c tio n s o n Affec tiv e Co m p u ti n g 1 0. 1 1 0 9/TAF FC. 2 0 2 2. 3 2 2 1 5 5 4 file

Emotion Recognition from EEG Network Connectivity Using Low-Dimensional Discriminant Analysis on Riemannian Manifolds

“…Euclidean space to Euclidean space methods have been most widely proposed and evaluated [15], [18]- [21]. The input data can be flexible such as EEG temporal time series and spectral features, which are usually treated as vectors or images that lie in the Euclidean space.…”

“…A typical machine learning method first extracts feature embedding from the input data and then classifies the feature embedding, both within the Euclidean space. For example, Zheng [15] treated differential entropy (DE) spectral features as vectors and a deep neural network to extract feature embedding for emotion recognition in the Euclidean space; Ding [18] treated multi-channel EEG temporal time series as images and then used a multi-scale convolution neural network (CNN) for feature extraction to realize emotion recognition again in the Euclidean space; Song [19] treated Euclidean EEG feature vectors as nodes on a graph and used the graph CNN for classification; Du [20], [21] used the more recent attention mechanisms and transformer architecture for end-to-end emotion recognition from Euclidean EEG temporal time series. However, the EEG network connectivity, as an important source of emotion-related neural information, has not been widely used in Euclidean space to Euclidean space methods because of its matrix data structure and Riemannian geometric structure.…”

“…Early studies have used traditional statistical learning methods such as linear discriminant analysis (LDA), support vector machine (SVM), and common spatial pattern (CSP) to extract channel-wise spectrotemporal EEG features and apply spatial filters in high-dimensional Euclidean space to realize emotion recognition [15]- [17]. More recent studies have used deep learning methods such as convolutional neural networks (CNN), recurrent neural networks (RNN), and graph neural networks (GNN) to recognize emotion directly from channelwise EEG features in an end-to-end manner, treating EEG features as high-dimensional vectors or images in Euclidean space [18]- [21]. On the other hand, recent neuroscience findings have shown that emotion states are not only encoded in spatial-spectral-temporal EEG features but also encoded in functional network connectivity such as the covariance and coherence among different brain signal channels [22], [23].…”