Spectral clustering is a method that employs the spectrum of a similarity matrix. The spectrum of a matrix represents the set of its eigenvalues, and a similarity matrix consists of similarity scores between each data point. It reduces the dimensionality of data before clustering. In other words, we can say that spectral clustering creates a graph of data points, and these points are mapped to a lower dimension and separated into clusters.
A similarity matrix converts data to conquer the lack of convexity in the distribution. For any dataset, the data points could be n-dimensional, and here could be m data points. From these m points, we can create a graph where the points are nodes and the edges are weighted with the similarity between points. A common way to define similarity is with a Gaussian kernel, which is a nonlinear function of Euclidean distance:
The distance of this function ranges from 0 to 1. The fact...
