With　time-varying　volatility　and　individual　differences,　EEG　signals　are　difficult　to　analyse.　The　recognition　performance　of　the　traditional　feature　extraction　is　lowered　due　of　the　difficulty　in　tracking　the　dynamic　changes　of　EEG.　In　this　paper　the　Common　Spatial　Subspace　Decomposition　(CSSD)　algorithm　was　improved(named　Improved-CSSD),　putting　forward　a　kind　feature　extraction　method　which　has　the　performance　of　adaptive　ability.　This　method　introducded　control　parameters,　which　added　the　training　samples　of　the　assistants　to　that　of　the　target　subject　in　some　way.　Finally,　based　on　the　data　of　the　international　BCI　competition　database,　some　simulation　experiments　were　conducted　by　recognizing　EEG　signals　by　Improved-CSSD　and　SVM.　Compared　with　the　traditional　CSSD,　classification　accuracy　was　increased　about　8.26%　by　Improved-CSSD.　The　result　showed　that　the　approach,　proposed　in　this　paper,　had　a　good　adaptability　and　a　low　time　loss.　©　2012　IEEE.