Imaging　sensors　are　widely　used　in　HCI　applications　to　capture　images　for　facial　expression　recognition.　The　proccess　involves　extraction　of　features　from　captured　images　and　use　of　machine　learning　algorithms　like　K-NN　classification　to　identify　the　specific　expression.　We　propose　here　a　facial　expression　recognition　system　based　on　non-negative　matrix　factorization　(NMF).　As　facial　parts　are　more　prominent　to　express　a　particular　facial　expression　rather　than　whole　faces　and　NMF　does　part　based　analysis,　we　are　interested　to　analyse　how　NMF　works　for　Facial　expression　Recognition.　We　benchmark　our　NMF　based　system　on　CK+　and　JAFFE　dataset.　We　get　a　significant　result.　In　addition　we　also　propose　WAPA　and　OEPA　based　NMF　for　this　application.　Our　proposed　WAPA　and　OEPA　is　actually　two　types　of　fusion　method　where　WAPA　counts　the　all　four　parts　of　facial　features　and　we　name　it　as　Weighted　All　Parts　Accumulation　(WAPA)　algorithm.　On　the　otherhand,　OEPA　counts　only　the　most　expressive　parts　for　each　expression　and　we　name　it　as　Optimal　Expression-specific　Parts　Accumulation　(OEPA).　The　experiment　shows　our　proposed　WAPA　and　OEPA　based　NMF　outperform　the　prevalent　NMF　method.　Copyright　©　2014　ACM.