DBSCAN clustering (Ester et al. 1996) is a non-hierarchical clustering method. Unlike K-means or K-medoids, the number of clusters does not need to be pre-defined. Any distance measure can be used. DBSCAN is a popular clustering method in the machine learning community. It does not always perform better than K-medoids but can trace out non-linear cluster boundaries.
The “MinPts” parameter is the minimum number of points required to form a dense region. By default it is set to the number of data dimensions dim*2, or N/4, whichever is smaller. It may be increased for large or noisy data sets.
The “Epsilon” parameter is the neighborhood radius. By default, it is set to 0.1 times the maximal pairwise distance in the data set. The value can be optimized using the K-distance graph.
In the output table, cluster assignment is indicated by a positive integer. The value -1 indicates a “noise point” that is not be assigned to any group.
K-distance graph: This graph shows the distance from each point to its MinPts-1 nearest neighbor, sorted from smallest to largest. You should look for a break in the slope of this curve. The corresponding “K-distance” may be used as a value for the parameter Epsilon.
Reference
Ester, M., Kriegel, H.-P., Sander, J., Xu, X. 1996. A density-based algorithm for discovering clusters in large spatial databases with noise. Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD-96), 226–231.