Chronic　mental　pressure　will　affect　one＇s　health　directly　by　bring　a　series　of　pathology　and　physiology　risks.　Effective　methods　of　evaluating　psychological　pressure　can　detect　and　assess　real-time　stress　states,　warning　people　to　pay　a　close　attention　to　their　health.　This　paper　is　focused　on　the　problem　of　individual　difference　in　the　stress　evaluation　process.　An　improved　support　vector　machine　(SVM)　evaluation　algorithm　in　automated　valuation　of　stress/non-stress　reaction　was　proposed　based　on　the　measurement　of　surface　myoelectrogram　signal.　The　algorithm　clustered　the　samples　and　gave　the　clustering　information　to　the　loss　function　of　SVM　to　achieve　training　samples　screening.　With　the　imbalance　problem　of　the　two　kinds　of　samples　after　screening,　the　weight　were　given　to　the　loss　function　to　reduce　classifier＇s　prediction　tendentiousness,　which　decreases　the　error　of　training　sample　and　makes　up　for　the　influence　made　by　the　unbalanced　samples.　The　improved　algorithm　increased　the　classification　accuracy　from　70.　34%　to　79.　31%,　while　algorithm　running　time　was　decreased　from　2026.5　s　to　541.3　s.　Experimental　results　show　that　the　algorithm　can　effectively　avoid　the　influence　resulting　from　the　individual　difference　on　stress　appraisal　effect.　Meanwhile　the　algorithm　reduces　the　computational　complexity.