A　face　recognition　method　is　proposed　using　large　margin　nearest　neighbor　(LMNN)　algorithm　to　improve　locally　linear　embedding　(LLE)　features　classification　performance.　LMNN　is　used　to　seek　a　linear　transformation　by　which　the　Euclidean　distance　in　the　transformation　space　could　equivalently　be　viewed　as　Mahalanobis　distances　in　the　original　space.　Hence,　the　kNN　classification　performance　could　be　improved.　Experiments　on　ORL　database　and　extended　YaleB　database　demonstrate　the　effectiveness　of　the　proposed　method.