Abstract: An improved net traffic identifier algorithm was proposed based on semi-supervised clustering. Symmetrical uncertainty was used to reduce the net flow attributes, and then kernel function was used to project the rest attributes to higher dimentional space. The train net flow was clustered in high dimentional space hierarchically. Smooth factor, sihouette coefficient and entropy controlled the cluster process to get a well result. Experiments show that the algorithm got flat clusters without any huge cluster and could identify most net flow even encrypted ones.