As　a　popular　leaning　algorithm,　Support　Vector　Machine　(SVM)　have　been　utilized　to　solve　the　problem　of　data　mining　and　knowledge　discovery.　However,　as　far　as　some　unbalanced　data　sets　of　multi-group　are　concerned,　the　classifier　model　trained　by　C-SVM　always　presents　some　imbalanced　error-rates　on　separating　samples.　Based　on　analysis　of　Lagrange　multiplier,　the　paper　brings　forward　some　novelty　concepts　including　the　outer　boundary　of　group,　Misleading-SV,　prediction-error-rate,　etc.　An　innovative　SVM　based　on　C-correction　is　formulated　and　a　method　for　correcting　slack　constant　C　is　designed.　On　the　target　of　winter　wheat　seed　geometric　feature　evaluation　for　quality　gradation,　the　research　team　constructs　some　testing　experiments　for　method　validation.　Analysis　of　accuracy　contour　suggests　the　proposal　scheme　is　able　to　effectively　separate　seeds　by　their　geometric　property　at　an　accuracy　of　96.5%.　In　parallel　with　some　known　congeneric　algorithms,　contrast　results　evidences　that　as　the　data　set　with　sparse　samples　is　considered,　the　method　for　correcting　slack　constant　can　elevate　the　general　separation　precision　of　classifier　prominently.