SSPF: a Simple and Scalable Parameter Free Clustering Method.

Centroid-based clustering methods, e.g. K-Means, are widely used due to fast computation and clear interpretation. In this paper, inspired by the convex clustering technique, we propose a simple but very scalable centroid-based approach, named SSPF, that clusters billions of data points within a few minutes. More importantly, it is parameter-free, which means no prior knowledge of K, i.e. the ground-truth number of clusters, is needed. We provide both GPU and CPU versions. Comprehensive experiments are conducted to evaluate both the speed and clustering quality of SSPE It shows that SSPF is much faster than other parameter-free approaches, and very competitive to the state-of-the-art implementations of parametric methods. We also discuss the limitations and mitigation to provide a full view of this method.