In this chapter, we discussed two algorithms that can easily solve non-convex clustering problems. The first one is called DBSCAN and is a simple algorithm that analyzes the differences between points surrounded by other samples and boundary samples. In this way, it can easily determine high-density areas (which become clusters) and low-density spaces between them. There are no assumptions about the shape or the number of clusters, so it's necessary to tune up the other parameters to generate the right number of clusters.
Spectral Clustering is a family of algorithms based on a measure of affinity among samples. It uses a classic method (such as K-means) on subspaces generated by the Laplacian of the affinity matrix. In this way, it's possible to exploit the power of many kernel functions to cluster non-convex datasets.
We have also discussed two online algorithms...
