Abstract: Cluster analysis is an important research domain of data mining. On the unsupervised condition, it is aimed at figuring out the class attributes of samples in a mixed data set automatically. For decades a certain amount of clustering algorithms have been proposed associated with different kinds of priori knowledge. However, there are still some knotty problems unsolved in clustering complex data sets, such as the unknown number and miscellaneous patterns of clusters, the unbalanced numbers of samples between clusters, and varied densities within clusters. These problems have become the difficult and emphatic points in the research nowadays. Facing these challenges, a novel clustering method is introduced. Based on the definition of local density and the intuition of ordered density in clusters, the new clustering method can find out natural partitions by self-adapted searching the boundaries of clusters. Furthermore, in the clustering process, it can overcome the straitened circumstances mentioned above, with avoiding noise disturbance and false classification. The clustering method is testified on 6 typical and markedly different data sets, and the results show that it has good feasibility and performance in the experiments. Compared with other classic clustering methods and an algorithm presented recently, in addition, the new clustering method outperforms them on 2 different evaluation indexes.

Key words: data mining, clustering, local density, descending search, self-adaption