In social networks, some data may come in the form of bipartite graphs, where properties of nodes are public while the associations between two nodes are private and should be protected. When publishing the above data, in order to protect privacy, we propose to adopt the idea generalizing the graphs to super-nodes and super-edges. We investigate the problem of how to preserve utility as much as possible and propose an approach to partition the nodes in the process of generalization. Our approach can give privacy guarantees against both static attacks and dynamic attacks, and at the same time effectively answer aggregate queries on published data.