Nowadays, recommender systems are playing an important role in online platforms. Beyond accuracy, diversity has been recognized as an important metric to evaluate recommendation performance, where re-ranking is one of the main tools to pursue diversity in the recommendation. However, existing re-ranking strategies generally introduce diversity at the expense of sacrificing accuracy performance, since the trade-off between accuracy and diversity is not trivial to achieve. In addition, complementary relationship exists extensively between items. Complementary items themselves belong to different categories but have a very large correlation. Therefore, we introduce a novel diversified re-ranking model named CRDR by using complementary relationships of items, which increase the positions of items in the recommendation list that are complementary to the items previously purchased by the user, so as to improve the diversity of recommendations while ensuring accuracy. Specifically, we first mine complementary relationships in co-purchase items with graph attention networks. Then, we add our complementary scores to the MMR model and use a greedy strategy to solve the problem. Extensive experiments on two different datasets demonstrate that our CRDR model achieve a balance between accuracy and diversity over traditional state-of-the-art methods.