Feature weighted dual random sampling cluster Ensemble

Cluster ensemble combines multiple partitions of a set of objects into a stable and robust one. To obtain good ensemble performance, base clusterings are required to take into account quality and diversity. Recently, in spite of some researches focus on consensus function to prove ensemble quality, how to produce high-diversity and high-quality base clusterings at general step without global screening remains an open problem. For high-dimensional data, the clustering algorithm suitable for common data sets is extremely inefficient and there is basically no cluster in high-dimensional space. To get around this conundrum, subspace cluster ensemble was proposed. At present, although the random subspace cluster methods show good diversity, the quality of base clustering remains to be improved. This paper proposes an improved algorithm of dual random subspace cluster ensemble methods to ensure the high diversity of the base clustering while improving the quality at general step without global feature screening. Our method reduces the irrelevant, uninformative features of the class structure at the generation step of random subspace, making the class structure more obvious. The experiment demonstrates the effectiveness of our method.