In this chapter, we have presented some of the most important clustering algorithms that can be employed to solve non-convex problems. Spectral clustering is a very popular technique that performs a projection of the dataset onto a new space where concave geometries become convex and a standard algorithm such as K-means can easily segment the data.
Conversely, mean shift and DBSCAN analyze the density of the dataset and try to split it so that all dense and connected regions are merged together to make up the clusters. In particular, DBSCAN is very efficient in very irregular contexts because it's based on local nearest neighbors sets that are concatenated until the separation overcomes a predefined threshold. In this way, the algorithm can solve many specific clustering problems with the only drawback being that it also yields a set of noise points that cannot be...
