Information　overload　makes　a　big　obstacle　for　multimedia　services.　To　alleviate　the　burden,　collaborative　filtering　has　been　actively　studied　in　the　recommendation　field　to　help　users　find　satisfactory　content.　However,　current　methods　fail　to　comprehensively　predict　users’　interactions　since　users’　interests　are　complex　and　multifaceted.　We　argue　that　the　prevalent　graph-based　models　pay　attention　to　shared　characters　among　neighboring　node　users　or　items　while　ignoring　their　specific　characters.　In　this　work,　we　investigate　common　and　special　characters　hidden　in　users’　interactions　with　the　help　of　semantic　knowledge　to　model　users’　interests.　We　propose　a　dual-channel　common　and　special　embedding-based　collaborative　filtering　model　(CSE)　for　personalized　recommendation.　CSE　adopts　a　divide-and-conquer　strategy　to　capture　commonality　and　semantic　speciality　with　graph　convolutional　learning　and　knowledge　translation.　Ultimately,　common　and　special　embeddings　of　users’　interests　participate　in　pairwise　user-item　matching　to　score　their　interaction　for　personalized　recommendation.　Extensive　experiments　on　two　real-world　datasets　demonstrate　the　rationality　of　the　proposed　CSE.　It　verifies　that　the　proposed　CSE　can　comprehensively　model　user　interests　with　common　and　special　characters　and　improve　recommendation　performance.　©　2022,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.