Most　existing　typical　semi-supervised　learning　algorithms　focused　on　the　results　of　learning　while　facing　the　conflict　on　constraints.　And　most　solutions　use　unsupervised　distance-based　methods　to　adjust　the　conflicting　constraints　on　the　information　by　recalculating　the　samples＇　distance.　This　paper　presents　a　constraint-based　semi-supervised　dimensionality　reduction　algorithm　with　conflict　detection,　called　CDSSDR,　which　uses　the　information　of　priori　constraints　to　adjust　the　contradictions　in　the　constraints.　It　avoids　the　use　of　unsupervised　methods　to　adjust　the　prior　knowledge.　©2010　IEEE.