The output D represents low-rank descriptor for LDSs and can be employed for the classification of EEG trails.
We extract features by the above LDSs model and get two feature matrices A and C.
Considering LR-LDSs methods generating A on Finite Grassmannian, unlike two feature matrices (A, C) by LDSs, Euclidean Distance and Mahalanobis Distance can describe the distance between two feature spaces of EEG trails after LR-LDS.
From the above sections, we propose three methods for EEG pattern recognition: LDSs, LR+CSP, and LR-LDSs.
In LDSs model, the value of a hidden parameter describing dimension of Riemannian feature space is closely related to final accuracy.
Then five methods including CSP, CSSP, LDSs, LR+CSP, and LR-LDSs are compared with each other.
LDSs related methods outperform CSP and CSSP due to their both spatial and temporal features extraction.
Therefore, we choose ThNN algorithms for CSP, CSSP, LDSs, LR+CSP, and LR-LDSs methods uniformly.
LDSs can overcome these problems by extracting both spatial and temporal features simultaneously to improve the classification performance.
Caption: Figure 1: The relationship between hidden parameter and accuracy for LDSs. We choose "al" and "av," which are the highest and lowest accuracy performance, respectively, to show the relationship between hidden parameter and accuracy.