These years much effort has been devoted to improving the accuracy or relevance of the recommendation system. Diversity, a crucial factor which measures the dissimilarity among the recommended items, received rather little scrutiny. Directly related to user satisfaction, diversification is usually taken into consideration after generating the candidate items. However, this decoupled design of diversification and candidate generation makes the whole system suboptimal. In this paper, we aim at pushing the diversification to the upstream in the representation learning stage, with the help of Graph Neural Networks(GNN). Although GNN based recommendation algorithms have shown great power in modeling complex collaborative filtering effect which makes the recommended items more relevant, how diversity change is ignored in those advanced works. We propose to perform rebalanced neighbor discovering, category-boosted negative sampling and adversarial learning with the guidance of item similarity. We conduct extensive experiments on real world e-commerce datasets. Experimental results verified the effectiveness of our proposed method on providing diverse contents. Further ablation studies validate that our proposed method could significantly alleviate the accuracy-diversity dilemma.