The traditional clustering algorithms mostly use simple one-dimensional distance measurement and overall optimization strategy for classification, one-dimensional measurement method can identify similarity matrix, but cannot keep the spatial location relationship between samples, because of the loss of data information between the sample which cause the clustering accuracy problem, this paper puts forward a graph clustering algorithm based on gravitational vector. This method combines the principle of gravity principle, uses the multi-dimensional vector instead of the traditional one-dimensional distance measurement method, and excavates the connection between the samples by establishing the local balance relationship, and finally achieves the purpose of improving the local information extraction and data distribution perception ability of the algorithm. Using the present algorithm in artificial datasets and UCI datasets can show that the present algorithm can cope well with multi-type, unbalanced, high complexity data and get good clustering results.