Multi-Prototype Clustering Algorithm Based on Density Peak Sampling

The popularity of k-means algorithm as a prototype-based clustering algorithm is due to its simplicity and efficiency. However, it uses only one prototype to model each class of data, which makes it less effective when dealing with non-convex unbalanced datasets. To overcome this issue, researchers have proposed multi-prototype clustering algorithms. However, it remains a challenging problem to determine the number of prototypes and estimate the final number of clusters. In response to this issue, we propose a novel multi-prototype clustering algorithm called MP-DPS, which is based on density peak sampling. First, we design an incremental sampling method that utilizes reconstruction criterion to automatically determine the number of prototypes. Then, we design a subcluster merging method based on the convex clustering model that does not require a priori k. The effectiveness and superiority of the MP-DPS algorithm are demonstrated by experimental results on both synthetic and real datasets.