The visualization of the associative relation between objects is mainly expressed by the edges of the graph. But a large number of edges will cause serious visual confusion due to the complexity of the associative relation between objects. Graph layout and edge bundling are both effective methods to solve the problem of visual confusion caused by complex edges. While the geo-location of some nodes has significant meaning, only edge bundling methods can be used to reduce the map load and reveal the potential association rules of graph. In the past, the edge bundling algorithms adjusted the position of the middle control points of the edge under the condition that the two end nodes of the edge were fixed. This would cause a large number of edges being gathered together, which would not only cause a secondary visual confusion, but also be difficult to reveal the potential rules of the graph at the node level. To solve this problem, this paper proposed a node-centered edge clustering and visualization algorithm. Firstly, the direction clustering algorithm was used to realize the clustering of direction edges. The direction clustering method proposed in this paper was about 13 times faster than K-means algorithm, and about 6 times faster than DBSCAN algorithm. Then, the interpolation of the control points was implemented for each edge. On this basis, FR model was used to prevent the occurrence of “excessive bending”. Finally, we adjusted the transparency of the edges so that the result of the visualization would be able to highlight the portion of the edge near the end nodes. The experimental results show that the value of the NCEB algorithm's map load (L) and the mid-point distance change (△d) were about half of the FDEB algorithm, which proved that the NCEB algorithm can move the binding position from the middle part of the edge to the node, thus not only solving the secondary visual confusion caused by traditional edge bundling methods, but also revealing the association rule and trend of the graph at the node level. The final distribution trend of the edge around the node was clear and readable, and the visualization result greatly reduced the map load, which effectively reduced visual errors and misunderstanding of information. The results of our experiments show that the proposed algorithm can reveal the potential associated trend of graphs at the node level and greatly reduce visual confusion.