Credit card payment has become one of the most commonly used consumption methods in modern society, yet risks of fraud transactions using credit cards also increased. Numerous methods have been proposed for credit card fraud detection during past decades. However, most of the existing frameworks mainly focus on directly processing structured data. While they lack inner relations between features of raw descriptions for credit owners, this could lead to information deficiency. Therefore, we proposed a graph-based semi-supervised fraud detection framework. In this work, the structured dataset is translated to graph format through the sample similarity in order to improve the effect of label propagation on the graph. We further adopt the GraphSAGE algorithm which has been demonstrated to show excellent performance on node classification tasks. Experimental results on the real-world dataset show that our graph-based model can outperform state-of-the-art baselines. We argue that our model could be extended to other classification tasks using structured data.