In　brain-computer　interface　(BCI)　systems　with　a　small　number　of　training　samples,　a　method　called　as　regularized　common　special　subspace　decomposition　(R-CSSD)　algorithm　was　proposed　to　solve　the　problems　such　as　low　stability　of　the　eigenvalues　and　poor　discriminative　ability　of　eigenvectors　in　electroencephalography　(EEG)　recognition　process.　In　R-CSSD,　regularization　was　introduced　based　on　the　traditional　common　special　subspace　decomposition　(CSSD)　algorithm.　The　presented　method　was　composed　of　three　steps:　First,　the　training　samples　of　the　specific　subject　could　be　effectively　combined　with　those　of　the　other　ancillary　subjects　by　two　regularization　parameters;　Second,　a　regularized　special　filter　was　built,　and　then　the　feature　information　of　the　specific　subject＇s　EEG　was　extracted;　Finally,　K-nearest　neighbor　(KNN)　algorithm　was　used　to　identify　motor　imagery　EEG.　Under　small-sample　condition,　the　experimental　results　show　that　R-CSSD　algorithm　not　only　can　effectively　improve　the　stability　of　the　eigenvalues　of　EEG,　but　also　can　produce　high　classification　accuracy　and　less　time　consumption.