OPTICS clustering — a reachability plot, then a cut
OPTICS orders the rows so that dense neighbors are adjacent in the order, and it records how far each row had to reach to join that order. Valleys in the reachability plot are cluster candidates. The xi cut looks for steep drops and rises; it does not commit to one eps up front. That is why one plot can show structure at more than one scale.
On scaled Iris, minSamples 5 and the xi cut produced 2 clusters with silhouette 0.657, the highest silhouette of the Iris runs, and adjusted Rand only 0.444 against species. The plot found a clean dense valley that is not the three-species partition. A high silhouette here means the valleys are compact, not that the species labels were recovered.
Read the plot before you trust the labels. If the reachability curve is flat, the table does not have density structure and the cut is arbitrary.
Settings that change this fit
- minSamples
- Smoothing of the reachability curve. Larger values suppress tiny valleys that are only a handful of rows.
- xi
- How steep a valley must be before it is extracted. Lower xi yields more, smaller clusters.