Item-based　Collaborative　Filtering　(ICF)　is　wildly　used　recommendation　method.　Under　the　assumption　that　a　user　tends　to　choose　an　item　similar　to　what　she　has　interacted　before,　ICF　uses　items　that　a　user　has　interacted　to　learn　a　user＇s　profile,　and　then　makes　recommendations　by　calculating　the　similarity　between　a　target　item　and　the　user＇s　profile.　Therefore,　the　method　of　learning　item　representations　from　data　is　crucial　in　ICF.　However,　most　existing　ICF　methods　only　consider　the　low-order　connectivity　between　a　target　item　and　a　target　user＇s　interacted　items.　The　consideration　limits　the　improvement　of　recommendation　qualities.　In　this　paper,　by　using　the　message-passing　idea　of　Graph　Convolution　Networks　(GCN),　we　propose　GCN-ICF　to　capture　the　high-order　connectivity　between　users　and　items　in　the　user-item　interaction　bipartite　graph　structure.　The　method　enriches　the　representation　of　each　item　through　its　neighbors　in　the　bipartite　graph　structure.　The　experimental　results　on　three　public　datasets　verify　the　highly　positive　effect　of　the　performance　of　the　GCN-ICF　method.