A new unsupervised clustering method based on outlier information

Traditional clustering algorithms such as CURE and ROCK require the user to provide the number of final clusters k, and outliers are treated as "noise" in the clustering process. By regarding outliers as valuable information, this paper takes a new perspective and complements with classical approaches. The proposed method integrates outlier identification with cluster number determination, leading to a more robust and truly unsupervised learning paradigm. To demonstrate its feasibility, two improved clustering algorithms CURED and As-ROCK are constructed based on CURE and ROCK. Empirical results demonstrate that these two novel algorithms not only can stop automatically, but also gain much in performance.