Tumor　gene　expression　data　has　the　characteristic　of　high　dimensionality　and　small　sample　size,　which　pose　a　rigorous　challenge　for　tumor　classification.　Since　not　all　the　genes　are　associated　with　tumor　phenotypes,　the　irrelevant　features　seriously　reduce　the　learning　performance.　It　is　necessary　to　select　relevant　features　from　the　original　data.　In　this　paper,　we　propose　a　new　filter　feature　selection　method　based　on　the　graph　embedding　framework　for　manifold　learning,　which　is　named　as　LLRFC　score.　The　relationship　between　sample　classes　and　features　is　considered　in　this　method.　But　the　selected　features　via　this　method　may　contain　some　redundancy.　Thus　it　is　improved　through　eliminating　redundancy　among　the　features.　The　improved　method　is　named　LLRFC　score+.　Several　other　feature　selection　approaches　are　used　to　compare　with　our　method　on　nine　public　tumor　gene　expression　datasets,　the　experimental　results　demonstrate　that　our　presented　method　is　quite　promising　and　valid　for　tumor　classification.　©　2016　IEEE.