To deal with the self-publication system’s huge text scale and speedy increasing, a right recommendation technology is helpful to realize the market value of “Long Tail” books. To deal with this issue, Directed Graph Clustering Algorithm is presented. Regarding clustering as the process of objects identifying, key words’ election depends on how much information they transfer in context. Moreover, to improve the efficiency of “Long Tail” texts’ clustering, a floating threshold and a sharing pool are set. Finally, experimental results comparing the K-Means algorithm prove that this clustering algorithm based on the directed graph is self-adaptive and effective.