Recommender systems on E-Connerce platforms track users' online behaviors and recommend rel-evant items according to each user's intersts  and needs. Bipartite graphs that capture both user/item feature and use-item interactions have been demon-strated to be higly effective for this purpose.Recently, graph neural network (GNN) has been successfully applied in representation of bipartite graphs in industrial recommender systems. Provid-ing individualized recommendation on adynamic platform with billions of users is extremely chal-lenging.A key observation is that the users of an online E-Commerce platform can be naturally clus-tered into a set of communities.We propose to slus-ter the users into a set of communities and make recommendations based on the information of the users in the community colletively.More specifi-cally, embeddings are assigned to the communities and the users information is decomposed into two parts, each of which captures the community-level generalizations and individualized preferences re-spectively. The community structure can be con-sidered as an enhancement to the GNN methods that are inherently flat and do not learn hierarchi-cal representations of graphs. The performance of the proposed algorithm is demonstrated on a public dataset and a world-leading E-Commerce company dataset.