Chapter 8 Spectral Clustering

This final chapter dealing with the treatment of classic clustering methods considers a totally different approach. Spectral Clustering is a graph partitioning method that can be interpreted as simultaneously implementing dimension reduction with cluster identification. It is designed to identify potentially non-convex groupings in the multi-attribute space, something the other cluster methods are not able to do. It has become one of the go-to methods in modern machine learning.

Here again, the focus will be only on those aspects of spectral clustering that differ from what was covered in the previous chapters.

To illustrate spectral clustering, the well-known spirals data set from the literature will be used, shown in Figure 8.1. It is included as one of the GeoDa sample data sets. The distinctive pattern consists of 300 observations on two non-intersecting spiral point clouds.

Figure 8.1: Spirals Data Set