S3LRR: A Unified Model for Joint Discriminative Subspace Identification and Semisupervised EEG Emotion Recognition

Emotion recognition from electroencephalogram (EEG) data has been a research spotlight in both academic and industrial communities, which lays a solid foundation to achieve harmonic human–machine interaction. However, most of the existing studies either directly performed classification on primary EEG features or employed a two-stage paradigm of “feature transformation plus classification” for emotion recognition. The former usually cannot obtain promising performance, while the latter inevitably breaks the connection between feature transformation and recognition. In this article, we propose a simple yet effective model named semisupervised sparse low-rank regression (S³LRR) to unify the discriminative subspace identification and semisupervised emotion recognition together. Specifically, S³LRR is formulated by decomposing the projection matrix in least square regression (LSR) into two factor matrices, which complete the discriminative subspace identification and connect the subspace EEG data representation with emotional states. Experimental studies on the benchmark SEED_V dataset show that the emotion recognition performance is greatly improved by the joint learning mechanism of S³LRR. Furthermore, S³LRR exhibits additional abilities in affective activation patterns exploration and EEG feature selection.