Robust common spatial patterns with sparsity

Robust and sparse modeling are two important issues in brain-computer interface systems. L1-norm-based common spatial patterns (CSP-L1) method is a recently developed technique that seeks robust spatial filters by using L1-norm-based dispersions. However, the spatial filters obtained are still dense, and thus lack interpretability. This paper presents a regularized version of CSP-L1 with sparsity, termed as sp-CSPL1. It produces sparse spatial filters, which eliminate redundant channels and retain meaningful EEG signals. The sparsity is induced by penalizing the objective function of CSP-L1 with the L1-norm. The sp-CSPL1 approach uses the L1-norm twice for inducing sparsity and defining dispersions simultaneously. The presented sp-CSPL1 algorithm is evaluated on two publicly available EEG data sets, on which it shows significant improvement in classification accuracy.