To　improve　expression　recognition　rates,　Kappa-based　contribution　degree　computing　approach　of　face　sub-area　in　recognizing　expression　is　proposed.　It　is　utilized　as　the　basis　of　deriving　weights　to　fuse　the　subspace　prediction　results.　The　normalized　face　emotion　images　are　partitioned　averagely　into　two　sub-regions　to　obtain　the　three　expression　subspaces　containing　the　upper　and　lower　half　face　parts　and　the　whole　face.　Each　part　is　represented　with　Gabor　feature.　Then,　three　classifiers:　SMO,　MLP　and　KNN　are　separately　used.　The　expression　prediction　results　are　counted　to　obtain　the　Kappa.　Experiments　are　done　on　the　CMU　and　JAFFE　two　expression　image　databases　and　results　show　that　Kappa　weighted　fusion　expression　recognition　approach　has　higher　recognition　accuracy.